`mkdocs build --strict` only catches references it cannot resolve. A bare
anchor such as `[limit][]` or `[vector search][search]` is matched by
autorefs against any heading on the site, so six of them silently linked
into the JavaScript reference instead. The relative links in
`permutation.py` and `remote/errors.py` pointed at in-page anchors and
paths that do not exist.
Targets that still exist here or in an imported inventory now use
mkdocstrings references; the guide pages deleted in #2770 use their
lancedb.com URLs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`mkdocs build` emitted 61 warnings on main, and rendering the previously
undocumented classes in this PR pushed that to 158. That backlog is what
blocks turning on strict mode (#3707), so clear it here rather than leave
it worse than we found it.
Most of it was one systematic false positive: griffe cannot see the
generated `__init__` of a pydantic dataclass, so every documented
parameter looked unknown. `warn_unknown_params` turns that check off.
The rest were real docstring bugs, in 15 docstrings:
* Prose trailing a `Parameters` section is read as parameter names, which
invented parameters called `The`, `you` and `To`. Moved into `Notes` or
the summary.
* numpydoc only reads a type when the colon has spaces around it. Where
the documented name is a pydantic attribute rather than a signature
parameter, griffe has no signature to fall back on and the type was
dropped. Affects nine embedding classes.
* `num_partitions, default sqrt(num_rows)` and friends parse as a list of
names, rendering a bogus `default` parameter.
* One parameter indented five spaces instead of four.
`nodejs/CONTRIBUTING.md` links to the repo-root CONTRIBUTING.md, which
does not resolve once typedoc copies the file into `docs/src/js/_media/`;
an absolute URL works from both places.
`mkdocs build --strict` now exits 0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four packages are now rendered by a single mkdocstrings directive each,
driven by the module's `__all__`, instead of a hand-maintained list of
symbols. These were where most of the drift was: 7 of 12 rerankers and
14 of 17 embedding functions had never been listed.
`lancedb.embeddings` had no `__all__`; without one mkdocstrings renders
no members at all for a re-export package, so one is added.
AGENTS.md gains a section describing how the reference page is wired up
and how to check a docs build locally, plus a step in the "adding a new
method on Table" checklist.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Python API reference page had drifted from the public API. Branch
management (`Branches` / `AsyncBranches`, which own `diff` and `merge`),
structured full-text query classes, take queries, blob helpers,
namespace connections, most rerankers and embedding functions, the
PyTorch dataloader, and several other public symbols were never listed,
so they did not appear in the rendered docs.
Also fixes docstring cross-references that pointed at guide pages which
have since moved off this site, and at unresolvable relative targets
(`[Table](Table)`, `[PyArrow Table](pyarrow.Table)`).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`test_read_consistency_interval` asserted that a table opened with a
100ms `read_consistency_interval` still read stale data immediately
after a concurrent write. The cache timestamp is set when the table is
opened and reads within the interval do not refresh it, so that
assertion only held if the intervening open/count/commit/count sequence
finished within 100ms of real wall-clock time. On a loaded CI runner it
did not: the TTL expired, `count_rows` refreshed synchronously, and the
test failed with `left: 1, right: 0`. This broke the Rust workflow on
`main` at 0bc08160 (a Python-only commit).
This pins the `background_cache` mock clock once `table2` has seeded its
cache, and advances it explicitly in place of `tokio::time::sleep`, so
the test controls when the interval elapses. Same approach as #3547.
With the clock pinned there is no real sleep left to be imprecise, so
the `cfg(not(target_os = "windows"))` guard is dropped and the test now
runs on Windows too.
Verified by inserting a stall before the write: 120ms reproduces the
original failure deterministically, and with this change the test still
passes with a 500ms stall.
Fixes#3712
## What
Expose custom FTS stop-word lists in the Python and TypeScript public
APIs, including their standalone tokenize helpers and remote index
creation.
This PR supports concrete string lists only. It does not add file or
LanceDB-table stop-word sources.
## Why
Rust already exposes Lance's custom stop-word list option. The Python
and TypeScript APIs did not pass it through, and local index details did
not retain the full tokenizer parameters needed by index-backed
tokenization after reopening a table.
## How
- Add `custom_stop_words` / `customStopWords` to the Python and
TypeScript FTS and tokenize options.
- Preserve `None` / `undefined`, empty lists, and list contents without
normalization.
- Load the persisted FTS segment parameters when returning local index
details.
- Serialize the concrete list in remote create-index requests.
- Keep Python and TypeScript tests thin; behavior, persistence, query
tokenization, and remote JSON coverage live primarily in Rust.
## Validation
- `cargo check --quiet --features remote --tests --examples`
- `cargo clippy --quiet --features remote --tests --examples`
- `cargo test --quiet --features remote --tests`
- Python extension rebuild with `uv` and `maturin`
- Targeted Python tests: 4 passed
- Python `ruff format --check` and `ruff check`
- TypeScript build, typecheck, Biome lint, generated docs, and targeted
tests
---------
Co-authored-by: Yang Cen <yangcen@Yangs-Mac-mini.local>
## Summary
- add a keyword-only `transform_parallelism` option to
`StreamingDataset`
- preserve CPU auto-detection by default and fall back to one worker
when unavailable
- apply the configured limit to both the transform executor and
concurrency semaphore
- document and test explicit, default, fallback, and invalid values
## Testing
- `uv run --extra tests --with torch pytest
python/tests/test_elastic_dataloader.py -q` (`136 passed`)
- `uvx ruff check python/lancedb/streaming.py
python/tests/test_elastic_dataloader.py`
- `uvx ruff format --check python/lancedb/streaming.py
python/tests/test_elastic_dataloader.py`
- `git diff --check origin/main...HEAD`
Closes#3695
Co-authored-by: buduoqiu <yaodong-shen@users.noreply.github.com>
Standard GitHub-hosted runners are free on public repos, so all Actions
spend here is on the `*-8x-*` / `4x` larger runners. Measured over 30
days at current (post-Jan-2026) larger-runner rates, that is ~$1,400/mo,
and `npm-publish` is ~70% of it.
## Changes
**Fat LTO was forcing builds onto large runners.** `[profile.release]`
in `.cargo/config.toml` sets `lto = "fat"` with `codegen-units = 1`,
which is single-threaded and the peak-memory step. The macOS
`npm-publish` build was 111 of its 113 minutes in one `napi build` step,
making it the critical path of the whole publish pipeline. The ThinLTO
override already applied to Windows now covers macOS too, and both
Windows builds move from `windows-2025-8x-x64` to the free standard
`windows-2025`.
**The npm-publish cargo cache never existed.** There are zero caches
with its key prefix. The key was static, so `actions/cache` (which only
writes on a miss) could never refresh it, and a multi-GB release
`target/` per target could never fit the repo's 10 GB budget anyway. Now
caches only the crate registry, keyed on `Cargo.lock`. The docker builds
also mounted `.cargo/registry/*` while the cache saved `.cargo-cache`,
so containers re-downloaded the registry every run.
**Cache eviction thrash.** Repo cache usage is 10.4 GB against GitHub's
10 GB cap, so every PR run evicted main's warm entries. `rust.yml` and
`nodejs.yml` now restore everywhere but only save from `main`.
**npm-publish moves to nightly + tags** instead of every push to main
(~90/month). The cross-compiled targets do need watching, so
`report-failure` now fires on scheduled runs, and dedupes onto an
existing open issue rather than filing one per night.
**rust.yml aarch64-pc-windows-msvc** cross-compiled its tests and then
skipped them, paying full codegen and link cost for a compile check.
`windows-11-arm` is now GA and free on public repos, so it builds and
tests natively. Its test step also passes `--target` — without it cargo
used `target/ci/` rather than `target/<triple>/ci/` and rebuilt the
entire dependency graph a second time.
**pypi-publish.yml had no concurrency group**, so force-pushes left a
~74 minute Windows job running.
## What is cost vs. wall-clock
| Change | Cost | Wall-clock |
|---|---|---|
| Windows npm-publish → free runners | **−$570/mo** | slower per job
(8→4 cores) |
| npm-publish nightly | **−$125/mo** | — |
| pypi-publish concurrency | small | — |
| macOS ThinLTO | $0 (already free) | **−~50 min** per release |
| rust aarch64 Windows native | $0 (already free) | **−~25 min** |
| rust `--target` on test step | $0 | large, avoids a second full build
|
| rust-cache `save-if` | small | faster via real cache hits |
## Risks
- The two Windows builds now have 4 cores instead of 8 and ~14 GB of
free disk. If they fail, it is most likely disk rather than memory;
fallback is `windows-2025-4x-x64`, which still halves that line.
- `windows-11-arm` has a thinner toolset (choco/vcpkg/protoc under
emulation) and this enables a test step that has never run, so it may
surface real aarch64 failures. That is the point, but it is the change
most likely to need iteration.
- ThinLTO applies to published macOS and Windows binaries, typically
within a few percent of fat LTO. Linux release builds are untouched.
## Follow-ups
- `python.yml` `pydantic1x` (37 min) and `Doctest` (33 min) each rebuild
the extension from source via `pip install -e .` with no Rust cache;
they should consume the wheel the `linux` job already builds. Worth
~$235/mo and ~70 min of compute per run. Separate PR.
- The three `ubuntu-2404-8x-x64` npm-publish builds (~$420/mo at the old
cadence) are the remaining large-runner spend;
`aarch64-unknown-linux-gnu` could run natively on free
`ubuntu-24.04-arm`. Worth doing after this lands so the ThinLTO change
can be validated first.
- The wheel composite actions declare `python-minor-version` as required
but never use it, and every caller omits it (actionlint warns).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
`list_versions()` against a remote table on a server that uses
lance-namespace was failing. The server was returning
`timestamp_millis`, while db-catalog deployments were returning
`timestamp`, and the client was only accepting `timestamp`. So, updated
the client to accept both. (assuming we're migrating over time;
eventually we can turn off the `timestamp` code path I suppose.)
## Summary
Lance can now plan multiple byte ranges for the same blob in one
`read_blob_ranges` operation, but LanceDB users currently cannot expose
a complete set of logical ranges to that planner.
This complements `BlobFile`: file-like consumers such as PyAV can
continue to discover ranges dynamically, while callers that already know
the ranges for a batch can submit them together.
## Motivating example
A training table may store a large video blob together with a small
application-level clip index:
```text
video: blob
clips: [{offset, length}, ...]
```
The caller can select the videos and clips for a batch, obtain their row
IDs from the query, and read all of the selected windows together:
```python
rows = (
table.search()
.select(["clips"])
.with_row_id(True)
.limit(64)
.to_arrow()
.to_pylist()
)
requests = []
for row in rows:
clip = sample_clip(row["clips"])
requests.append(
(row["_rowid"], clip["offset"], clip["length"])
)
chunks = table.fetch_blob_ranges("video", requests)
```
Here, `_rowid` comes from the LanceDB query, while `offset` and `length`
come from the application's clip index and are relative to that row's
video blob. The caller describes only the logical reads; Lance still
handles validation, source grouping, coalescing, scheduling, and byte
backpressure.
Lance v10.0.0-beta.5 returns one logical result per blob selector or
range request and explicitly distinguishes null blobs from valid empty
values. LanceDB consumes that aligned result contract directly and only
adds a cardinality check for unresolved row IDs.
This PR exposes batched blob-range reads on local Rust and Python
tables. Results preserve request identity, duplicates, null slots, and
valid empty ranges while allowing Lance to execute the physical reads
out of order. Scheduler buffer sizing remains an internal Lance concern,
so the LanceDB API does not expose `io_buffer_size`.
Cloud tables continue to report this operation as unsupported until
there is a corresponding remote API.
SELECT COUNT(*) FROM t WHERE <predicate> — and any query that plans an
empty-projection scan — panics the executing query task:
InvalidArgumentError("must either specify a row count or at least one
column")
Root cause
MetadataEraserExec wraps every LanceDB table scan to strip schema-level
metadata, rebuilding each batch in execute():
RecordBatch::try_new(schema.clone(), batch.columns().to_vec()).unwrap()
RecordBatch::try_new infers the row count from the columns. COUNT(*)
with a filter is planned with an empty projection, so the scan emits
zero-column batches — there are no columns to infer a length from,
try_new returns Err, and the .unwrap() panics.
(This is specific to the empty-projection case: COUNT(*) with no filter
is answered from statistics and never scans, and COUNT(<col>) projects a
column — both already work.)
Updates the Rust workspace Lance dependencies and Java lance-core
dependency to v10.0.0-beta.5. No compatibility fixes were required;
full-workspace Clippy passes with warnings denied. Lance tag:
https://github.com/lance-format/lance/releases/tag/v10.0.0-beta.5
## What
MemWAL LSM **read** support. When a table has an LSM write spec
(`set_lsm_write_spec`), `merge_insert` upserts live in the MemWAL
active/frozen memtables and flushed SSTables until an external
compaction merges them into the base table, so a normal scan returns
**stale** data. This routes reads through Lance's `LsmScanner` so
queries also surface that in-flight data, deduplicated by primary key
(newest generation wins).
## How
- Adds a **`use_lsm: Option<bool>`** query flag, symmetric with the
`merge_insert` flag:
- **unset** — auto-route through the LSM scanner when the table carries
a write spec
- **`use_lsm(true)`** — force the LSM path; error if there is no spec
- **`use_lsm(false)`** — read the base table only (the escape hatch)
- Plain scan, single-column full-text search, and single-vector ANN all
run through one `LsmScanner` (assembled from on-disk shard manifests
plus the cached writer's in-memory memtables), so a `where` predicate is
honored as a **prefilter** uniformly — including for vector search.
- **Compaction-aware snapshots:** an SSTable generation is dropped only
once it is both compacted into the base table and covered by the arm's
base-index catch-up (`index_catchup`); plain scans use the compaction
watermark alone.
- Query shapes the scanner cannot honor hard-error with guidance to set
`use_lsm(false)`: hybrid, multi/binary vectors, `with_row_id`,
reranking, `order_by`, dynamic/Substrait projection or filters,
`distance_range`, `use_index(false)`, postfilter, take-by-row-id/offset,
reads from a time-traveled version, and an unmaintained or ambiguous
(multiple) FTS/vector index. Namespace-pushdown queries fall back to
local execution when a spec is present; WAL-only writers are handled.
- Exposed across the Rust core and the Python (`use_lsm`) and TypeScript
(`useLsm`) bindings, including `TakeQuery`.
Rebased from Lance `7.2.0-beta.3` to `10.0.0-beta.3`.
Python was versioned and tagged separately from the Rust, Java, and
Node.js SDKs, and had drifted three minor versions ahead (0.36 vs 0.33).
Users had no way to tell which Python version corresponded to which Rust
or Node release, and the gap had no meaning behind it.
This unifies the two tracks so there is one version and one tag for all
four SDKs.
## Version
The shared version is set to `0.37.0-beta.0`. Python continues its own
sequence (highest published: 0.36 → 0.37) while Rust, Java, and Node.js
jump 0.33 → 0.37 to meet it. Picking Python's next minor means Python
users see no discontinuity at all, and only the other SDKs skip forward.
Note that `main` trails the `release/v0.32` branch on both lines (main
is at 0.32.0-beta.3 / 0.35.0-beta.3; the release branch carries
0.33.0-beta.0 / 0.36.0-beta.0), so 0.37 is chosen to clear the highest
tag on either branch. Every index stays monotonic:
| index | publishes | last published | next |
|---|---|---|---|
| PyPI | stable only | 0.34.0 | 0.37.0 |
| Fury | previews | 0.36.0b0 | 0.37.0-beta.1 |
| npm | both | 0.33.0-beta.0 | 0.37.0-beta.1 |
| crates.io | stable only | 0.31.0 | 0.37.0 |
| Maven | both | 0.33.0-beta.0 | 0.37.0-beta.1 |
A one-time jump for three SDKs, versus explaining the offset
indefinitely.
## Mechanism
* `python/.bumpversion.toml` is removed. `python/Cargo.toml` — the
source of the Python package version, since `pyproject.toml` declares
`dynamic = ["version"]` — becomes a tracked file of the root config. Its
`cargo update -p lancedb-python` pre-commit hook is dropped as
redundant: `ci/update_lockfiles.sh` already refreshes every workspace
member version in `Cargo.lock`.
* `pypi-publish.yml` triggers on `v*` instead of `python-v*`, so one tag
releases all four packages. `ci/bump_version.sh` and
`make-release-commit.yml` lose their now-dead tag-prefix and
per-language plumbing, including the `python` / `other` dispatch inputs.
* The two byte-identical GH release jobs in `npm-publish.yml` and
`pypi-publish.yml` are replaced by a single `gh-release.yml`. One
release per tag, named `LanceDB vX.Y.Z`, instead of separate "Python
LanceDB" and "Node/Rust LanceDB" releases for the same commit.
The trade-off: there is no longer a way to ship a Python-only patch
without also releasing crates.io, Maven, and npm. That is the cost of
making drift structurally impossible.
## Beta releases marked "Latest" (#3666)
Both GH release jobs used:
```yaml
prerelease: ${{ contains('beta', github.ref) }}
```
The arguments are reversed. `contains(search, item)` asks whether
*`search`* contains *`item`*, so this evaluated "does the literal string
`'beta'` contain `refs/tags/python-v0.35.0-beta.2`?" — always `false`.
Every beta was published as a full release, and GitHub awards "Latest"
to the newest non-prerelease.
The new workflow derives the flag from the parsed version rather than
the raw ref, and sets `make_latest` explicitly:
```yaml
prerelease: ${{ steps.extract_version.outputs.prerelease }}
make_latest: ${{ steps.extract_version.outputs.prerelease == 'false' }}
```
npm was never affected (`--tag preview` uses correct bash), and PyPI
already excludes pre-releases from resolution.
This only fixes releases published from here on. Already-published betas
need a one-time backfill:
```shell
gh api --paginate /repos/lancedb/lancedb/releases \
--jq '.[] | select(.prerelease == false) | select(.tag_name | test("beta")) | .id' \
| xargs -I{} gh api -X PATCH /repos/lancedb/lancedb/releases/{} -F prerelease=true
```
## Verification
Ran `ci/bump_version.sh` end-to-end against this branch with the release
tooling installed:
* `preview` → tags `v0.37.0-beta.1` (previous tag `v0.33.0-beta.0`
detected, `pre_n` bump)
* `stable` → tags `v0.37.0`
* Both paths update `.bumpversion.toml`, `rust/lancedb/Cargo.toml`,
`nodejs/Cargo.toml`, `python/Cargo.toml`, `nodejs/package.json`, the 7
`nodejs/npm/*/package.json` files, both Java poms, and
`docs/src/java/java.md` together
* `check_breaking_changes.py` resolves the last stable as `v0.31.0`, so
the minor-version gate passes
All five touched workflows parse as valid YAML and the pre-commit hooks
pass.
## Notes for review
* This targets `main` only, so it takes effect at the next
release-branch cut. The in-flight `release/v0.32` branch still carries
`v0.33.0-beta.0` / `python-v0.36.0-beta.0`; if we want the imminent
stable to be 0.37.0, this needs to be applied there too.
* Historical `python-v*` tags are left alone. The changelog builder
scans `^v`, which does not match them, so the first unified release's
notes will compute `fromTag` from the Rust/Node line only — a one-time
gap in the Python-side changelog.
* Pre-existing and not addressed here: `ci/update_lockfiles.sh --amend`
amends the commit that `bump-my-version` has already tagged, so the
lockfile update lands outside the tag on stable releases.
Fixes#3666
## What changed
- add `block_size` to Python FTS configuration and the deprecated
local/remote helpers
- add `blockSize` to the TypeScript FTS options and propagate it through
the NAPI binding
- serialize the value as `block_size` for remote index creation
- document the existing Rust builder API and generate the TypeScript API
reference
- add local, remote, metadata, search, and invalid-value regression
coverage
## Why
Lance supports configuring the number of documents per compressed FTS
posting block, but LanceDB's Python and TypeScript APIs did not expose
the setting. This made the experimental FTS V3 layout unavailable
through those clients and allowed the value to be dropped before index
creation.
## How it works
The default remains `128`. Supported values are `128` and `256`;
selecting `256` uses the experimental FTS V3 format. Invalid values are
rejected by the Lance builder and surfaced as Python or JavaScript
errors.
## Validation
- `cargo check --quiet --features remote --tests --examples`
- `cargo +1.94.0 clippy --quiet --features remote --tests --examples --
-D warnings`
- targeted Rust local and remote index tests
- Rust doctests: 34 passed
- Python Ruff checks, doctest, and targeted local/remote tests: 5 passed
- TypeScript build, Biome lint, generated docs, and targeted Jest tests:
9 passed
- `git diff --check`
## Limitations
The Java client remains unchanged because its external remote REST model
does not currently expose `block_size`.
Co-authored-by: Yang Cen <yangcen@Yangs-Mac-mini.local>
## What
`AnswerdotaiRerankers(return_score="all").rerank_hybrid(...)` (and
`ColbertReranker`, which subclasses it without overriding
`rerank_hybrid`) raises:
```
pyarrow.lib.ArrowInvalid: Invalid sort key column: No match for FieldRef.Name(_relevance_score) in _rowid: int64 ...
```
## Why
```python
combined_results = self.merge_results(vector_results, fts_results)
combined_results = self._rerank(combined_results, query)
if self.score == "relevance":
combined_results = self._keep_relevance_score(combined_results)
elif self.score == "all":
combined_results = self._merge_and_keep_scores(vector_results, fts_results)
```
When `score == "all"`, `combined_results` is unconditionally overwritten
by `_merge_and_keep_scores(vector_results, fts_results)` **after**
`_rerank()` already computed and appended `_relevance_score` —
discarding it. The following `sort_by("_relevance_score", ...)` then has
nothing to sort on.
Every sibling reranker that supports `return_score="all"`
(`cross_encoder`, `openai`, `cohere`, `jinaai`, `voyageai`, `watsonx`)
instead calls `_merge_and_keep_scores()` **before** `_rerank()`. This
file is the one place the ordering got inverted when `"all"` support was
added (#2509) — a copy/paste inconsistency across the six files that PR
touched. Fix mirrors the pattern already used (and tested) by the other
five rerankers.
Also drops the now-stale `"Only 'relevance' is supported for now"`
docstring line on both classes, left over from before `"all"` support
existed.
## Testing
Added `test_answerdotai_reranker_return_all`, mirroring the existing
`test_cross_encoder_reranker_return_all`. Verified locally with the real
built Rust extension: red (reproduces the exact `ArrowInvalid` above) →
green, using the actual `rerank_hybrid`/`_rerank`/`base.py` code path
with the model call mocked out — my local environment's
`rerankers==0.10.0` fails to load the real ColBERT model against the
available `transformers` version (`AttributeError: 'ColBERTModel' object
has no attribute 'all_tied_weights_keys'`), which I confirmed also
breaks the **pre-existing**, unmodified
`test_colbert_reranker`/`test_answerdotai_reranker` baseline tests
identically — an unrelated local dependency-version issue, not a
regression from this change. `ruff check`/`ruff format` clean; full
`test_rerankers.py` run: 9 passed / 8 skipped / 3 failed (the 3 failures
are exactly those two pre-existing tests plus my new one, all failing at
model-loading time for the same unrelated reason before reaching the
changed code).
---
Disclosure: this PR was drafted with AI assistance (Claude); I reviewed,
tested, and take responsibility for the change.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
## Summary
This updates the Java API reference to close the documentation gaps that
can be fixed from the current Java source and generated namespace API.
The patch adds an empty table example, shows how to wrap returned Arrow
IPC query bytes in a reusable `ArrowFileReader` helper, and documents
the Java index operations that are currently exposed by the namespace
client: vector indexes, scalar indexes, full text search indexes, and
listing indexes.
## Issue Links
Fixes https://github.com/lancedb/docs/issues/157
Fixes https://github.com/lancedb/docs/issues/160
Partially addresses https://github.com/lancedb/docs/issues/159 by
documenting the index parameters currently exposed by Java. The
requested `num_partitions` example is still blocked because
`CreateTableIndexRequest` does not expose IVF training parameters yet.
Not included: https://github.com/lancedb/docs/issues/158. The current
Java docs and source remain remote namespace oriented, so local DB
connection documentation should wait until the Java local DB API is
available and can be verified.
## Validation
- Built the Java core module with OpenJDK 17:
`./mvnw -pl lancedb-core -am -DskipTests compile`
- Checked the Markdown diff:
`git diff --check -- docs/src/java/java.md`
The Java build succeeds. It still reports pre-existing checkstyle
warnings in the namespace client builder, but the Maven build is green.
## Summary
- reconstruct foreign Arrow Map schemas from their single sanitized
entries field
- reject malformed Map types with anything other than one child
- preserve the complete Map schema and `keysSorted` value through
empty-table creation and IPC round trips across Arrow 15–18
## Testing
- `./node_modules/.bin/jest --runInBand __test__/arrow.test.ts
__test__/sanitize.test.ts`
- `pnpm lint`
- `pnpm build`
- `pnpm run docs`
Fixes#2337
## What
- Replace legacy model names in `WatsonxEmbeddings` with the current
supported set:
- `ibm/granite-embedding-278m-multilingual` (new default, 768-dim)
- `ibm/slate-125m-english-rtrvr-v2` (768-dim)
- `ibm/slate-30m-english-rtrvr-v2` (384-dim)
- `intfloat/multilingual-e5-large` (1024-dim)
- `sentence-transformers/all-minilm-l6-v2` (384-dim)
- Add `space_id` field — mutually exclusive with `project_id`, mirrors
the
existing pattern in `WatsonxReranker`
- `project_id` / `space_id` resolution now falls back to
`WATSONX_PROJECT_ID` /
`WATSONX_SPACE_ID` env vars; exactly one must be supplied
## Why
The previously hardcoded models (`ibm/slate-125m-english-rtrvr`,
`sentence-transformers/all-minilm-l12-v2`) are legacy and no longer
listed as
supported by the watsonx.ai platform. `space_id` scoping was already
supported
by `WatsonxReranker` but was missing from the embeddings counterpart.
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
Updates the Rust workspace Lance dependencies and Java lance-core from
v9.1.0-beta.5 to v9.1.0-beta.7, including the generated Cargo lockfile.
No LanceDB compatibility changes were required for this release. See the
[Lance v9.1.0-beta.7
release](https://github.com/lance-format/lance/releases/tag/v9.1.0-beta.7).
This PR adds some support for `diff` / `merge` in the remote client as
for local tables we stay `NotSupported` until
https://github.com/lance-format/lance/issues/7263.
This wires the two review-and-land calls against the remote REST API:
- `POST /v1/table/{id}/branches/diff`
- `POST /v1/table/{id}/branches/merge`
Rust gets typed results (`BranchDiff`, `MergeBranchResult`). Python
returns the wire JSON, same shape as the REST response.
Merge here means promoting a branch's added columns onto `main`.
### Behavior
- Remote only. Local raises `NotSupported`.
- A rejected merge is not an exception. HTTP 409 still returns `Ok` / a
dict with `status="rejected"` and blockers in `diff.mergeBlockers`.
- Unknown blocker / status codes parse as `Unknown` so a newer server
does not break older clients.
- `MergePreview` tolerates missing fields for the same reason.
- Merge requests are not retried. 409 is final and carries the body you
need.
### Example
```python
table = db.open_table("images")
table.branches.create("exp")
exp = table.branches.checkout("exp")
exp.add_columns({"tag": "cast('draft' as string)"})
diff = table.branches.diff("exp")
preview = table.branches.merge("exp", dry_run=True)
result = table.branches.merge("exp", dry_run=False)
if result["status"] == "merged":
print("landed at", result["mainVersionAfter"])
elif result["status"] == "rejected":
print(result["diff"]["mergeBlockers"])
```
### Testing
cargo test -p lancedb --features remote diff_branch
cargo test -p lancedb --features remote merge_branch
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
## Problem
On the remote (LanceDB Cloud) write path, each write partition is
uploaded as a **single** `/insert?upload_id=...` request that stays open
until the whole partition has been streamed and the server has written
it to object storage. For large bulk ingests a partition can be many GB,
so a single request can run longer than the client read timeout (default
300s), surfacing as:
```
lancedb.remote.errors.HttpError: operation timed out
```
The server already supports staging **multiple** parts under one
`upload_id` (each `/insert` writes a separate transaction that
`complete` merges atomically), but the client never used that — it sent
one part per partition.
## Change
Split each partition into multiple parts of at most
`max_bytes_per_request` (Arrow IPC, LZ4-compressed) bytes, each uploaded
as its own `/insert?upload_id=...&upload_part_id=...` request. This
bounds how long any single request stays open, independent of total data
size or write parallelism.
Key properties:
- **Still streamed, not buffered.** Each part's body is driven through a
bounded channel while the request is in flight (`futures::join!` of a
producer + the send), so peak memory stays at a couple of batches per
partition regardless of the part size. Backpressure from a
slow/throttled server still propagates upstream.
- **Correct part accounting.** An empty partition still sends exactly
one (schema-only) part so `complete` has a transaction to commit; a size
cut landing exactly on the end of input does not emit a trailing empty
part.
- **Multipart only.** The single-request (non-multipart) path is
unchanged.
## Config
New `ClientConfig::max_bytes_per_request: Option<usize>`, also settable
via the `LANCE_CLIENT_MAX_BYTES_PER_REQUEST` environment variable.
**Default 1 GiB** (`Some(0)` disables splitting → one request per
partition). Python users pick up the default/env automatically through
the remote client.
## Tests
- `test_multipart_chunked_splits_into_parts`: a 1-byte budget puts each
batch in its own part → N requests, each carrying the shared `upload_id`
and a distinct `upload_part_id`.
- `test_multipart_single_part_when_under_budget`: a large budget keeps
the partition in a single request.
- Verified end-to-end against a live remote table: a forced-chunked
multipart add (many parts) assembles to the correct row count.
Related to ENT-1883.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
Fixes#1653.
`infer_vector_column_name` in `util.py` could silently return `None`
when `query is None` and `query_type` is not `"fts"` or `"hybrid"`. This
`None` then propagated into downstream code, causing a cryptic
`TypeError: expected bytes, NoneType found` rather than a clear error
message.
## Changes
- **Removes the no-op `try/except Exception as e: raise e`** around
`inf_vector_column_query` (it was catching and immediately re-raising
without adding any value)
- - **Adds a `None` guard** after the inference block: if
`vector_column_name` is still `None` at this point, raise a clear
`ValueError` pointing the user to pass `vector_column_name` explicitly
## Before / After
**Before:** cryptic `TypeError: expected bytes, NoneType found` deep in
schema lookup code
**After:**
```
ValueError: No vector column found in the schema. Please specify the vector column name explicitly via the `vector_column_name` parameter.
```
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
Some additions to our lancedb skill to enable agents to use the jobs
methods that we recently added. Eval tests (below, with and without
these additions to the skill) suggest that they're helping, mostly to
find the right method calls. These are a little unusual because they
require REST server connection, they're not yet implemented in the SDKs.
```
┌─────────────────────┬───────────┬────────────┬─────────────┬──────────┬───────────┬──────────┬───────────┐
│ eval │ grade w/o │ grade with │ improvement │ time w/o │ time with │ cost w/o │ cost with │
├─────────────────────┼───────────┼────────────┼─────────────┼──────────┼───────────┼──────────┼───────────┤
│ 8-list-running-jobs │ 2.5/3 │ 3/3 │ +0.5 │ 123s │ 29s │ $0.58 │ $0.18 │
├─────────────────────┼───────────┼────────────┼─────────────┼──────────┼───────────┼──────────┼───────────┤
│ 9-describe-job │ 1/5 │ 5/5 │ +4.0 │ 159s │ 52s │ $0.62 │ $0.25 │
├─────────────────────┼───────────┼────────────┼─────────────┼──────────┼───────────┼──────────┼───────────┤
│ 10-cancel-job │ 3/3 │ 3/3 │ +0.0 │ 99s │ 35s │ $0.55 │ $0.21 │
├─────────────────────┼───────────┼────────────┼─────────────┼──────────┼───────────┼──────────┼───────────┤
│ TOTAL │ 6.5/11 │ 11/11 │ +4.5 │ 381s │ 116s │ $1.75 │ $0.65 │
└─────────────────────┴───────────┴────────────┴─────────────┴──────────┴───────────┴──────────┴───────────┘
```
Failure reasons are because the agent didn't know the right method to
call, spent all its turns guessing REST calls, tried to inspect lancedb
code, but didn't find the answer in here.
## Problem
`table.add(dataset)` with a `pyarrow.dataset.Dataset` OOMs the client
during bulk ingestion of wide rows (e.g. embedding columns), even
against a remote table where the upload itself is streaming.
The cause is in `to_scannable`: a `Dataset` is scanned with pyarrow's
default scanner settings (`batch_size=131072` rows,
`batch_readahead=16`, `fragment_readahead=4`). pyarrow's internal
threads prefetch that read-ahead window independently of LanceDB's
backpressure, so for wide rows a large fraction of the dataset is held
in memory. On the remote path this is then multiplied across the
multipart write partitions (one in-flight batch per partition, up to
CPU-core count).
Reproduced on a 10 GB / 1.55M-row dataset with two 768-dim float32
embeddings: peak client RSS ~11.7 GB for the scan alone (6.8 GB after
consuming a *single* batch), ~15.4 GB for the full remote `add()`.
## Fix
`to_scannable` now sizes the scanner from an estimate of bytes-per-row
derived from the schema:
- **Narrow datasets keep pyarrow's defaults** (empty scanner kwargs) —
no throughput regression. The bound only engages above ~410 bytes/row.
- **Wide rows** get a smaller `batch_size` (~16 MiB/batch) and reduced
read-ahead (`batch_readahead=2`, `fragment_readahead=1`) so peak
in-flight memory stays near a ~1 GiB budget. Read-ahead (not just batch
size) has to drop, because pyarrow pins whole row-group buffers.
On the 10 GB dataset this drops peak client RSS to ~1.4 GB, and it stays
flat as the dataset grows. The `Dataset`/`LanceDataset` scannables
remain rescannable (retry-safe).
## Also: expose `write_parallelism` on `add()`
`AddDataBuilder::write_parallelism` already existed in Rust but was not
exposed in Python. This PR forwards it through the async, sync, and
remote `add()` methods, so users can cap the number of parallel write
partitions (each buffers data in flight) to trade throughput for memory
on large uploads.
## Tests
- `test_scannable.py`: bytes-per-row estimation; narrow → defaults; wide
→ bounded; `Dataset` reader streams bounded batches and stays
rescannable.
- `test_table.py`: `write_parallelism` on sync and async `add()`, and
that `write_parallelism=0` is rejected.
Fixes ENT-1883
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Moves the skill from .agents/skills/lancedb to
plugins/lancedb/skills/lancedb, as recommended by codex and claude.
Install path now is:
### Codex/ChatGPT app
Codex: Plugins -> "Create" arrow -> Add plugin marketplace
search for lancedb plugin and install
### Codex CLI
```
codex plugin marketplace add lancedb/lancedb
codex plugin install lancedb@lancedb # name@marketplace
```
### Claude app
Settings -> Plugins -> Add -> Add marketplace
search for lancedb plugin and install
### Claude CLI
```
claude plugin marketplace add lancedb/lancedb
claude plugin install lancedb@lancedb
```
Here's how it looks on ChatGPT/Codex app:
(the main icon has light and dark modes; the smaller one on the skill
doesn't so I made it gray 🤷 )
<img width="764" height="560" alt="Screenshot 2026-07-16 at 2 49 24 PM"
src="https://github.com/user-attachments/assets/b82cda16-3392-4740-ac47-b2f187cb2655"
/>
Hi, and thank you for LanceDB.
Small CI supply-chain hardening. In `make-release-commit.yml`, the
release job checks out with `LANCEDB_RELEASE_TOKEN` (a push-capable PAT)
and its final step pushes the version tag using a third-party action
pinned to a **mutable branch**:
```yaml
- name: Push new version tag
uses: ad-m/github-push-action@master
with:
github_token: ${{ secrets.LANCEDB_RELEASE_TOKEN }}
```
`@master` can move after review; whatever it points at then runs with
that release token in scope. This PR pins it to the commit behind the
current release (`v1.3.0` → `881a6320…`), keeping the version visible as
a comment. Behavior today is unchanged.
For transparency: I used AI assistance to spot and draft this; I
verified the workflow and resolved the SHA myself.
Tracks #3324. On x86_64 CPUs without AVX2 (Sandy Bridge / Ivy Bridge /
Westmere on Intel; Bulldozer / Piledriver / Steamroller on AMD), `import
lancedb` SIGILLs because the wheel bakes AVX2 + FMA into every compiled
function. Per [westonpace's
review](https://github.com/lancedb/lancedb/issues/3324#issuecomment-4328944354),
the default `lancedb` wheel stays fast; pre-Haswell users get a
separately-published `lancedb-compat` wheel.
## Summary
- Adds a `lancedb-compat` matrix entry to `pypi-publish.yml` that builds
with `RUSTFLAGS="-C target-cpu=x86-64-v2"` (Nehalem-class baseline).
Same Python API (`import lancedb` works) — files install to the same
namespace, so the two wheels conflict at install time and users pick
one. Same pattern as `psycopg2` / `psycopg2-binary` and `tensorflow` /
`tensorflow-cpu`.
- Generalizes `build_linux_wheel` and `upload_wheel` composites with
optional `package-name` and `rustflags` inputs (defaults preserve the
existing 4 `lancedb` matrix entries verbatim).
- Documents the choice in `python/README.md`: `pip install
lancedb-compat` for pre-Haswell hosts.
The default `.cargo/config.toml` baseline is unchanged.
## Sequencing
1. ~~lance-format/lance#6630 merges → runtime SIMD dispatch lands in
lance.~~ **Done — merged.**
2. lancedb's lance dep is bumped to a release that includes it (separate
PR / normal cadence).
3. This PR's `lancedb-compat` wheel build path starts producing a wheel
that runs on pre-Haswell hardware. **Maintainer setup**: register
`lancedb-compat` on PyPI and configure trusted publishing.
## Verified end-to-end on Sandy Bridge Xeon E5-2609
Verification was done locally against a fork-pinned lance dep that
includes the runtime dispatch implementation, using the same
`RUSTFLAGS="-C target-cpu=x86-64-v2"` flags this PR uses in CI:
```
$ RUSTFLAGS="-C target-cpu=x86-64-v2" maturin build --release
$ pip install ./target/wheels/lancedb-*.whl
$ python verify.py
PASS: import + simd dispatch + table create + vector search all work.
```
Pre-fix on the same CPU (default `pip install lancedb`): `Illegal
instruction (core dumped)`. Full reproducer (deps + clone + build +
verification):
https://gist.github.com/tobocop2/2e341358b55c143527416edfdb1e37df.
Fork-internal verification PR with the dep bump and full logs:
[`tobocop2/lancedb#2`](https://github.com/tobocop2/lancedb/pull/2).
## Benchmarks — no regressions on modern CPUs from the lance-side change
These are the numbers I ran for the lance PR, confirming the runtime
dispatch doesn't slow down the default (`target-cpu=haswell`) wheel that
existing users install. Criterion, one machine, one session, base → PR,
no `RUSTFLAGS` override. Full methodology, null experiments, and logs:
[lance-format/lance#6630 benchmark
comment](https://github.com/lance-format/lance/pull/6630#issuecomment-4933063394)
and the [logs
gist](https://gist.github.com/tobocop2/3c6d0f449cbd736aa2501f89a7fe56a2).
| benchmark | EPYC 7B13 (`avx2`, `fma`, no `avx512f`) | Xeon Cascade
Lake (`avx512f`) |
|---|---|---|
| `Cosine(f32, scalar)` *(control)* | +0.04% | +0.09% |
| `Cosine(f64, scalar)` | −0.34% | −1.94% |
| `Cosine(u8, SIMD)` | +2.30% | +3.63% |
| `Dot(f16, SIMD)` | −0.58% | +0.61% |
| `Dot(f32, SIMD)` | +0.34% | **−6.08%** |
| `Dot(f32, arrow_arity)` | +0.02% | −0.00% |
| `L2(f32, scalar)` | −0.10% | −0.02% |
| `L2(f32, simd)` (dim 1024) | +2.63% | −0.53% |
| **`L2(simd,f32x8)` (dim 8)** | **−45.9%** | **−25.1%** |
| `L2(u8, SIMD)` | +0.42% | −3.11% |
| `NormL2(f32, SIMD)` | −1.02% | −4.17% |
| `NormL2(f64, SIMD)` | +3.51% | −0.58% |
Nothing regresses beyond the noise floor. Dim 8 — the PQ sub-vector
width — improves 25–46%.
---
To be transparent: this isn't my domain of expertise and the lance-side
implementation is AI-generated. I verified it works end-to-end on the
failing hardware. Happy to roll in feedback.
Routes local sync child-namespace operations through the Rust-backed
connection instead of the Python namespace-client fallback.
Also keeps lazy namespace-client construction for table-to-Lance
conversion and preserves public namespace error mappings.
Validated locally with ruff format/check and targeted namespace pytest.
BREAKING CHANGE: splits generated by the permutation data loader will
not be the same, due to a change in hash function.
Updates the Lance dependencies and Java lance-core to
[v9.0.0-rc.1](https://github.com/lance-format/lance/releases/tag/v9.0.0-rc.1).
Includes the required DataFusion 54 and Lance file-reader compatibility
updates.
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds client-side support for analyze_plan distributed metrics modes
across Rust, Python, and TypeScript clients. Defaults to aggregate for
backward compatibility and sends the remote distributed_metrics
parameter only when a non-default mode is requested.
Fixes#3174
Also fixes#3645
Empty record batches now append correctly typed empty embedding arrays
without invoking embedding providers. This avoids OpenAI requests with
an invalid empty input while preserving source-column validation and
the non-empty execution paths.
As a small cleanup, the single- and multi-embedding code paths now share
a single upfront lookup of their source columns ("input_columns")
instead
of each path looking them up independently. Also moves `lance-testing`
from regular dependencies to dev-dependencies where it belongs.
Tests run:
- `cargo fmt --all -- --check`
- `cargo test --quiet -p lancedb --lib
empty_batch_skips_embedding_functions`
- `cargo test --quiet -p lancedb --lib
empty_batch_still_validates_source_column`
- `cargo test --quiet -p lancedb --lib
test_create_empty_table_with_embeddings`
- `cargo check --quiet -p lancedb --features remote --tests --examples`
- `cargo clippy --quiet -p lancedb --features remote --tests --examples`
- `cargo test --quiet -p lancedb --lib`
- `cargo test --quiet --features remote --tests`
## Summary
Fix `on_bad_vectors="fill"` so it replaces only invalid or missing
vector values instead of replacing the entire vector row.
Fixes#3026.
## Reasoning
The old Python sanitizer detected whether a vector row was bad at row
granularity. For `fill`, it then used that row-level flag to replace the
whole vector with `[fill_value] * dim`. That meant an input like `[1.0,
NaN, 3.0]` became `[0.0, 0.0, 0.0]`, even though the documented and more
useful behavior is to preserve valid values and fill only the bad
element.
I checked whether this should be a Rust-side fix so TypeScript users
would benefit too. Today, Rust core exposes `NaNVectorBehavior::{Error,
Keep}` for rejecting or keeping NaN vectors, while the Python
`on_bad_vectors` API (`error`, `drop`, `fill`, `null`) is implemented in
the Python ingestion sanitizer before data reaches Rust. TypeScript does
not expose the Python `on_bad_vectors="fill"` behavior today. Moving
this exact behavior to Rust would be a broader cross-language API
change, so this PR keeps the fix scoped to the currently affected Python
API.
## What changed
- Added a small helper that fills bad vector rows by preserving valid
elements, replacing NaN elements with `fill_value`, truncating vectors
longer than the expected dimension, and padding short vectors with
`fill_value`.
- Kept the existing fast path unchanged: the helper only runs after bad
vectors are detected and `on_bad_vectors="fill"` is selected.
- Updated sanitizer and table tests to assert element-wise NaN
replacement and short-vector padding for both `create_table` and `add`.
## Validation
- `uv run ruff format .`
- `uv run ruff check .`
- `cd python && uv run --no-sync pytest
python/tests/test_util.py::test_handle_bad_vectors_jagged
python/tests/test_util.py::test_handle_bad_vectors_nan
python/tests/test_table.py::test_create_with_nans
python/tests/test_table.py::test_add_with_nans -vv`
Targeted pytest result: `10 passed`.
## Why this fix is Python-side (and not Rust)
The problematic behavior lives in Python’s `on_bad_vectors` sanitizer,
before data is handed off to Rust. Rust currently only exposes
`NaNVectorBehavior::{Error, Keep}` for add operations, while Python has
the richer `on_bad_vectors={"error","drop","fill","null"}` API.
TypeScript does not currently expose the Python-style fill behavior, so
moving this exact fix into Rust would require designing a broader
cross-language bad-vector handling API.
This PR keeps the change scoped to the existing affected surface:
Python’s `on_bad_vectors="fill"` path. This way, Python users
immediately benefit.
## What the new agent skill covers
We want to help users _easily_ write LanceDB pipelines to bring their
data in from other places, no matter whether they use LanceDB OSS or
Enterprise.
The `lancedb` set of skills contains guidance for agents on the
following:
- Distinguishes local and remote table capabilities.
- Promotes bounded reads using `select()` and `limit()`.
- Prevents accidental full-table materialization.
- Documents correct Python sync/async scan APIs.
- Recommends validated Python schemas and batched ingestion.
- Provides indexing, query-tuning, diagnostics, and maintenance
guidance.
- Documents the Enterprise table-name cache issue: avoid immediately
reusing a dropped or overwritten table name; write to a fresh name and
rename after propagation.
- Adds Python and TypeScript API, pattern, and performance references.
- Adds a heuristic scanner for potentially unsafe Python and TypeScript
materialization patterns.
This change only adds agent documentation and tooling: no LanceDB
runtime code, Rust code, SDK APIs, dependencies, or CI configuration are
modified.
## Context
The LanceDB agent skill was accidentally pushed directly to `main` in
`8ea78e3fbcb26718112ab4ddec55a91804b869d3`, bypassing the normal review
workflow. That commit was reverted on `main` by `c12a6dce` so the
protected branch is back to its prior content.
## Summary
- add table-level FTS query tokenization returning token text and
position
- use the native index tokenizer for local tables and remote index
metadata for remote tables
- expose sync and async Python table wrappers with focused coverage
`Dataset::index_statistics()` loads index files and does meaningful CPU
work to serialize low-level info. Most fields
`NativeTable::index_stats()` needs are available from manifest metadata
via `Dataset::describe_indices()`, which is much cheaper.
`NativeTable::index_stats()` now:
- Calls `describe_indices()` filtered by name; returns `Ok(None)` if no
match.
- Parses `distance_type` from `description.details()` JSON (the
`VectorIndexDetails` proto stored in the manifest by recent Lance
versions).
- Falls back to `index_statistics()` only for vector indices where
`details()` returns no `distance_type` — this handles older Lance
datasets that didn't write `VectorIndexDetails`.
- `Unknown` index types (e.g. Lance's internal `FragReuseIndex`) are
explicitly filtered out of `list_indices` rather than erroring.
## Test plan
- [x] `test_create_scalar_index` — asserts `index_type`,
`distance_type`, and `num_unindexed_rows > 0` after adding rows
post-index
- [x] `test_create_fm_index`, `test_create_bitmap_index`,
`test_create_label_list_index` — added `index_stats` assertions
- [x] IvfPq, IvfHnswPq, IvfHnswSq, IvfHnswFlat tests assert
`distance_type == Some(L2)`
- [x] `test_list_indices_skip_frag_reuse` — FragReuseIndex is filtered
by the Unknown guard in `list_indices`
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
- serialize sync phrase queries consistently for execution and query
plans
- restore the documented no-argument hybrid `phrase_query()` behavior
- keep reranker input as the original user text without mutating the
builder
Fixes#3653.
## Testing
- `python/.venv/bin/python -m pytest <8 focused test nodes> -q` (`8
passed`)
- `python/.venv/bin/python -m ruff format --check
python/python/lancedb/query.py python/python/tests/test_fts.py
python/python/tests/test_hybrid_query.py`
- `python/.venv/bin/python -m ruff check .`
- `git diff --check origin/main...HEAD`
The complete hybrid module and the real native FTS phrase test were not
completed
in the current PyO3 runtime environment: both stalled in the native
`lancedb.connect()` fixture and were interrupted without an assertion
failure.
The CODEOWNERS file added in #3312 automatically requests reviewers on
every PR — the `*` default owner routes all changes to two reviewers.
This is mostly noise for contributors, and we prefer a single requested
reviewer per PR.
Remove the file.
Reverts #3312.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Carrying over from #2915, this patch introduces:
* Single-API call batching support for Gemini embeddings (up to 100 at a
time, the API limit)
* A versioned user agent header for Gemini API calls
* Support for [variable embedding dimension
size](https://ai.google.dev/gemini-api/docs/embeddings#control-embedding-size)
(Gemini is MRL trained)
## Summary
- preserve explicit `0.0` distance bounds in synchronous hybrid search
- distinguish omitted `None` endpoints from zero-valued endpoints when
configuring the vector child query
- add a public end-to-end regression test for a zero upper bound
## Testing
- `cd python && uv run --extra tests pytest
python/tests/test_hybrid_query.py -q`
- `uv run --project python ruff format --check
python/python/lancedb/query.py python/python/tests/test_hybrid_query.py`
- `uv run --project python ruff check .`
Fixes#3651
description: Branch management for LanceDB tables via the REST API. Use this skill whenever someone wants to create, delete, list, or switch branches on a LanceDB table — or needs to make sure a write (metadata update, index build, etc.) lands on a specific branch instead of main. Invoke it even without the word "branch" if context makes clear they want an experimental copy of a table, want to isolate changes, or want to confirm a mutation didn't touch main. Covers: branches/list, branches/create, branches/delete, and passing "branch" in describe/update_field_metadata/create_index to target a non-main version.
---
## Goal
Manage branches on a LanceDB table: list what exists, create new ones, delete stale ones, and direct read/write operations at a specific branch without touching main.
## Step 0: Establish the connection
Use the `lancedb-connect` skill to resolve the base URL and auth headers (`x-api-key`, `x-lancedb-database`). Skip this only if the connection is already known from the current conversation.
All examples below use `{base_url}` — substitute the resolved endpoint and include the auth headers on every request.
## The branch model (important)
LanceDB branches are named snapshots that diverge from the table's current state at creation time. There is **no checkout command** — you never switch the whole table to a branch. Instead, you **pass `"branch": "<name>"` in the request body** of any operation to target that branch. Omitting the key (or sending an empty body) always targets main.
`branches/list` returns only non-main branches. Main always exists and is not listed.
description: Column metadata authoring for LanceDB tables via the REST API. This skill is required for tasks like writing field descriptions, setting tags on columns (field_type, model, project_id, version), classifying columns as embeddings vs labels vs eval metrics, or grouping versioned columns into logical families — because it has the API integration needed to read the schema and persist metadata back. Invoke whenever someone wants to document, annotate, tag, or classify what their table columns ARE. Trigger even without an explicit "LanceDB" mention, as long as the context is column-level documentation or tagging for an ML or vector database table.
metadata:
short-description: Write column descriptions, tags, and logical groupings to a LanceDB table
---
## Overview
This skill authors column-level metadata for a LanceDB table. It connects to a LanceDB deployment over its REST API, inspects the table schema, generates appropriate metadata, and writes it back.
## Step 0: Establish the connection
Use the `lancedb-connect` skill (invoke it via the Skill tool) to resolve the base URL and auth headers (`x-api-key`, `x-lancedb-database`) for whichever deployment the user is working against — enterprise/self-hosted or a local dev server. Skip it only if the connection details are already established in the conversation.
All examples below use `{base_url}` — substitute the resolved endpoint and include the resolved headers on every request.
## Metadata keys
All metadata uses namespaced keys:
| Key | Purpose | Example value |
|-----|---------|---------------|
| `lancedb:description` | Human-readable explanation of what the column contains | `"CLIP ViT-L/14 image embedding, L2-normalized (768-dim)"` |
| `lancedb:tag:<name>` | Flexible key-value tag; the suffix names the tag category | `lancedb:tag:field_type: "embedding"`, `lancedb:tag:model: "clip"`, `lancedb:tag:project_id: "foo"` |
| `lancedb:logical-column` | Logical group/family this column belongs to | `"clip_features"` |
Tags are open-ended — use whatever key suffix and value make sense given the user's intent. The tag suffix should describe *what is being classified* (e.g., `field_type`, `model`, `project_id`) and the value describes *how*.
## Step 1: Resolve the table identifier
You need:
- **Table name** (required) — e.g., `my_table` or `my_namespace.my_table`
- **Database name** — ask if not provided and not inferable from context; it goes in the `x-lancedb-database` header, never in the URL path
The table identifier in the URL path is typically `table_name` for a top-level table, or `namespace$table_name` if the table lives in a namespace. The API accepts a `delimiter` query parameter to parse compound identifiers (default `$`).
## Step 2: Describe the table
```http
POST{base_url}/v1/table/{table_id}/describe
Content-Type:application/json
{}
```
The response contains `schema.fields` — an array of field objects:
description: Resolve how to connect to a LanceDB deployment over the REST API — figure out the base URL, API key, and database header. Use this before making any REST requests to a LanceDB table, whenever the endpoint or auth setup is not already known. Also useful on its own when someone asks how to connect, authenticate, or curl their LanceDB instance.
metadata:
short-description: Resolve the base URL and auth headers for a LanceDB deployment
---
## Goal
Produce two things every REST request needs:
1.**Base URL** — the endpoint
2.**Headers** — `x-api-key`, and usually `x-lancedb-database`
## Resolution steps
1. If the user already gave a URL and API key (or said which environment they're working against), use that.
2. Otherwise, look for credentials already available in the environment:
- Env vars like `LANCEDB_URI` / `LANCEDB_HOST` / `LANCEDB_API_KEY`
- A LanceDB endpoint already running or port-forwarded locally (the REST default port is 2333, i.e. `http://localhost:2333`)
3. If you didn't find both pieces, ask the user directly: **"What's your LanceDB endpoint's URL, and what's your API key?"** Also ask which database to use if it isn't obvious. Don't guess or probe further — the user knows their deployment.
## Validating the connection
Make a cheap authenticated request and check the status:
"description":"Write, review, debug, and document LanceDB pipelines in Python and TypeScript that work across local LanceDB OSS tables and remote LanceDB Enterprise/Cloud tables.",
description:"Override [project] name in python/pyproject.toml (e.g. 'lancedb-compat'). Default keeps 'lancedb'."
required:false
default:"lancedb"
rustflags:
description:"RUSTFLAGS for the build container, as a single whitespace-free token (e.g. '-Ctarget-cpu=x86-64-v2'). Empty leaves RUSTFLAGS unset, keeping the defaults from .cargo/config.toml."
required:false
default:""
runs:
using:"composite"
steps:
@@ -27,6 +35,18 @@ runs:
ARM_BUILD:${{ inputs.arm-build }}
run:|
echo "ARM BUILD: $ARM_BUILD"
- name:Patch package name for variant build
if:${{ inputs.package-name != 'lancedb' }}
shell:bash
env:
PACKAGE_NAME:${{ inputs.package-name }}
run:|
# Swap the [project] name so this build produces e.g. lancedb-compat
# wheels. The package still installs files under the lancedb/
# namespace -- import lancedb still works after pip install.
sed -i.bak 's/^name = "lancedb"$/name = "'"$PACKAGE_NAME"'"/' python/pyproject.toml
"description":"Write, review, debug, and document LanceDB pipelines in Python and TypeScript that work across local LanceDB OSS tables and remote LanceDB Enterprise/Cloud tables, with idiomatic query/search patterns and performance defaults for ingestion, indexing, filtering, and diagnostics.",
"description":"Codex plugin for building LanceDB pipelines in Python and TypeScript.",
"author":{
"name":"LanceDB"
},
"keywords":[
"lancedb",
"vector-search",
"full-text-search",
"hybrid-search",
"python",
"typescript",
"pipelines"
],
"skills":"./skills/",
"interface":{
"displayName":"LanceDB",
"shortDescription":"Build LanceDB pipelines in Python and TypeScript.",
"longDescription":"Write, review, debug, and document LanceDB pipelines in Python and TypeScript that work across local LanceDB OSS tables and remote LanceDB Enterprise/Cloud tables, with idiomatic query/search patterns and performance defaults for ingestion, indexing, filtering, and diagnostics.",
"developerName":"LanceDB",
"websiteURL":"https://www.lancedb.com",
"category":"Developer Tools",
"capabilities":[
"Developer Tools"
],
"defaultPrompt":"Create a LanceDB table, embed sample text, and run a vector search.",
description: Use when writing, reviewing, debugging, or documenting LanceDB pipelines in Python or TypeScript, especially code that should work across local LanceDB OSS tables and remote LanceDB Enterprise/Cloud tables. Helps avoid non-portable full-table materialization, choose idiomatic query/search patterns, apply LanceDB performance defaults for ingestion, indexing, filtering, and diagnostics, and resolve connections to the remote server for Enterprise-only operations such as jobs.
---
# Building LanceDB Pipelines
Use this skill to produce LanceDB pipelines that are portable between local and remote tables (for LanceDB Enterprise/Cloud) and idiomatic for the selected SDK.
## LanceDB Table Modes
LanceDB has two common execution modes:
- **Local table**: embedded, open source, in-process LanceDB. The client opens data from a local path or object storage URI and executes queries in the application process.
- **Remote table**: LanceDB Enterprise/Cloud table opened through a `db://...` URI. The data may be very large, commonly backed by object storage, and queried through a remote service.
Do NOT assume local-only table helpers exist on remote tables. If the user asks for LanceDB Enterprise, Cloud, `db://...`, production remote access, or a remote table, focus on the remote table path: use `search()` / `query()`, keep reads bounded with `select()` and `limit()`, and avoid table-level full materialization APIs.
## Workflow
1. Identify the SDK: Python, TypeScript, or both.
2. Identify the table mode: local/embedded OSS, remote Enterprise/Cloud, or portable across both. If the user says "LanceDB Enterprise", choose the remote table path. If the task involves jobs in any way (listing, inspecting, creating, or canceling jobs), it is always the remote path and requires a remote server connection — see "Connecting to the LanceDB remote server" below before doing anything else.
3. Read the matching language branch before writing or changing code:
- Remote server connection resolution (jobs, raw REST): `references/remote_connect.md`
- Job operations REST API (list/describe/cancel/query_events): `references/remote_jobs.md`
4. Start with `patterns.md` for the selected SDK. Read `api_reference.md` when choosing method names or return collectors. Read `performance.md` when the task involves ingestion, indexing, filtering, query tuning, diagnostics, or large datasets. Read `column_metadata.md` when the task is documenting, tagging, classifying, or grouping table columns (field descriptions, `lancedb:tag:*` tags, logical column families). Read `branch_ops.md` when the task involves branch lifecycle (list/create/delete), writing to a non-main branch, or verifying a change stayed off main. Read `remote_connect.md` when the task involves jobs or direct REST access to an Enterprise deployment, and `remote_jobs.md` for the job REST methods themselves (list, describe, cancel, query_events).
5. For Python schemas, favor Pydantic models and validate records before writing. Use PyArrow schemas when Arrow-native, streaming, or highly dynamic data makes them materially better suited.
6. Prefer `search()` or `query()` builders with explicit `select()` and `limit()` for reads.
7. Avoid table-level full materialization in remote or portable code. This is the main local-vs-remote read pitfall.
8. After a successful embedded OSS ingestion, call `table.optimize()`. Do not call it for Enterprise/Cloud; remote maintenance is automatic.
9. For remote Enterprise/Cloud writes, never drop-then-reuse or `mode="overwrite"` the same table name — see "Enterprise: never drop-then-reuse the same table name" below. This is the main local-vs-remote write pitfall.
10. If reviewing an existing file or repo, run `scripts/check_materialization.py` on the relevant paths and inspect each finding before editing.
11. Cross-check unfamiliar or non-trivial API claims against the source tree instead of relying on memory.
## Core Portability Rule
Do not write code that assumes a local table API will exist on a remote table. Remote tables can be very large, so whole-table materialization helpers are intentionally unavailable or unsafe.
This does **not** mean result conversion is forbidden. Bounded query/search result collection is normal:
The unsafe pattern is table-level or unbounded collection, plus local-only dataset escape hatches in remote code:
- Python: `table.to_pandas()`, `table.to_arrow()`, `table.to_polars()`; `table.to_lance()` is local/OSS-only dataset access, not materialization
- TypeScript: `await table.toArrow()`, `await table.query().toArray()` without `limit()`
## Enterprise: never drop-then-reuse the same table name
LanceDB Enterprise/Cloud splits a **control plane** (DDL: create/drop/rename) from a **data plane** (query nodes that serve reads). Query nodes cache the resolved dataset for a table name for up to `table_cache_ttl` — **default 300 seconds (5 minutes)**. After you drop or overwrite a table, the control plane updates immediately but the data plane keeps serving the *old* dataset until that cache entry expires. During the window the two planes disagree.
The failure this causes: you `drop_table("t")` then immediately `create_table("t", ...)` (or `create_table("t", ..., mode="overwrite")`). The DDL returns success, but every query against `t` returns **`500 Internal Server Error`** (the query node resolves the stale/deleted dataset), and a fresh `describe` may still show the *old* schema/version. It looks like your write silently failed; it didn't — the name is cached.
**`mode="overwrite"` has the same problem** — it is a drop+create of the same name under the hood.
Rules for portable Enterprise ingestion:
1. **Never reuse a table name you just dropped/overwrote within the cache TTL.** Do not use `mode="overwrite"` to replace an existing Enterprise table in place.
2. To (re)load data, **write to a fresh table name** (e.g. `<table>_v2`, or a run-stamped suffix). A brand-new name has no cached data-plane entry, so writes and reads work immediately.
3. Before creating, `list_tables()` and **fail loudly if the name already exists** rather than overwriting — prompt for a new name.
4. To land on a specific final name that is currently occupied by an old table: drop the old table, **wait out the TTL (~5 min), then `rename_table(fresh_name, final_name)`**. Renaming onto a name whose old dataset is still cached hits the same race, so the wait is mandatory. `rename_table` is a supported control-plane op.
5. When you hand a table name back to a human, tell them which step still needs the propagation wait (usually: "the old `t` was dropped; run the rename in ~5 minutes").
This is Enterprise/Cloud-specific. Local/OSS tables have no separate data plane, so `mode="overwrite"` and immediate same-name reuse are fine there.
## Connecting to the LanceDB remote server
LanceDB Enterprise/Cloud deployments are served by a server implementing the lance-namespace OpenAPI spec (<https://github.com/lance-format/lance-namespace/blob/main/docs/src/spec.yaml>). Every remote (`db://...`) connection talks to such a server, and some operations exist only there. In particular, **all operations around jobs (listing, inspecting, creating, or canceling jobs) run server-side** — there is no local/OSS equivalent. Before any job work, or any direct REST call to an Enterprise deployment, read `references/remote_connect.md` to resolve the base URL, credentials, and database header and to validate the connection. Then use the four job REST methods documented in `references/remote_jobs.md` (list, describe, cancel, query_events).
## Script
Run the scanner when reviewing or modifying an existing codebase:
Manage branches on a LanceDB table: list what exists, create new ones, delete stale ones, and direct read/write operations at a specific branch without touching main. Use for branch lifecycle tasks, experimental/isolated table versions, targeting an operation at a non-main branch, or confirming a mutation did not affect main.
Works on local/OSS and remote Enterprise/Cloud tables, except merging a branch into main, which is Enterprise-only.
## The branch model (important)
Branches are isolated, writable lines of history forked from another branch (or a specific version). Writes on a branch never affect `main`.
There is **no global "switch branch" state** — you never repoint the whole table at a branch. Instead, **operations are scoped by which table handle you use**:
- The handle you got from `open_table(name)` / `openTable(name)` targets `main`.
- `branches.create(...)` and `branches.checkout(...)` return a **new table handle scoped to that branch**. Every read/write on that handle (add, update, `update_field_metadata`, `create_index`, search, …) lands on the branch.
- The original main handle is unaffected — keep it around to verify isolation.
`branches.list()` returns only non-main branches. Main always exists and is not listed.
## Python
`table.branches` is a property returning the branch manager; `table.current_branch()` tells you what a handle is scoped to (`None` = main).
```python
table = db.open_table("products") # scoped to main
# list — dict of name -> metadata (parent_branch, parent_version, ...); {} = only main
table.branches.list()
# create: forks from main by default and returns a handle scoped to the new branch
exp = table.branches.create("experiment-reindex")
exp = table.branches.create("exp2", from_ref="main", from_version=None) # optional fork point
# checkout an existing branch -> branch-scoped handle
wip = table.branches.checkout("wip-branch")
# with version= it pins to that version (read-only detached view); omit to track latest, writable
# operate on the branch simply by using its handle
assert b"lancedb:description" in (wip.schema.field("category").metadata or {})
assert b"lancedb:description" not in (table.schema.field("category").metadata or {}) # main untouched
```
Two handles on the same branch see each other's writes (e.g. `table.branches.create("exp")` and `db.open_table(name, branch="exp")`); main stays isolated.
## Merging a branch into main (Enterprise only)
Merge is available through the SDKs (`table.branches.merge(...)`) on **Enterprise tables only** — it is not supported on Cloud or local/OSS tables, which raise `NotSupported`.
`merge` takes the branch to merge **from** and a `dry_run` flag. Both the SDK method and the underlying REST endpoint **actually merge by default** (`dry_run=False`); pass `dry_run=True` to only preview. A rejected merge is **not an exception** — it returns a result with `status="rejected"` rather than raising, so inspect the return value. Use `branches.diff(from_branch)` to inspect a branch's pending diff without attempting a merge.
```python
exp = "experiment-reindex"
# preview only — returns status="ready" if it would merge cleanly
preview = table.branches.merge(exp, dry_run=True)
# actually merge (default)
result = table.branches.merge(exp)
if result["status"] == "merged":
print("landed at", result["mainVersionAfter"])
elif result["status"] == "rejected":
print(result["diff"]["mergeBlockers"]) # why it was refused
The result is the wire JSON, containing `status` (`ready` on a passing dry run, `merged` on success, `rejected` when refused — also `notImplemented`/`unknown`), the branch `diff` (including `mergeBlockers` explaining any rejection), a `preview` of the columns that would be promoted, and — after a real merge — `mainVersionAfter`.
### Merge preconditions
Merge only **promotes newly added columns** onto main; it does not replay arbitrary commits. Practically, a branch is mergeable only if it has **exactly one commit since it was created, and that commit added a column**. The merge is rejected (`status: "rejected"`, with `mergeBlockers` set) if:
- the branch was forked from another branch rather than directly from main
- main has advanced since the branch was forked
- the branch's rows changed since the fork (row counts must match main exactly)
- the branch removed columns or changed a column's type/nullability
- the branch added no columns (index-only changes are not merged)
### Adding a column in a single commit
Because the branch must contain just one column-adding commit, add the column with its values in one operation rather than add-then-backfill:
1. **SQL transformation** — `add_columns` with a SQL expression computed from existing columns, so the column lands populated in one commit.
2. **Precompute the values** — compute the column's values externally, then add the fully-populated column in a single operation (e.g. via `merge_insert`/`add_columns` with the data ready).
3. **Lance-format-level data evolution (pylance)** — use Lance's data evolution with backfill, documented at <https://lance.org/guide/data_evolution/#with-data-backfill>.
| Create from a fork point | `table.branches.create(name, from_ref=..., from_version=...)` | `await branches.create(name, fromRef, fromVersion)` |
| Get a branch handle | `table.branches.checkout(name)` or `db.open_table(t, branch=name)` | `await branches.checkout(name)` or `await db.openTable(t, { branch: name })` |
| Pin to a branch version (read-only) | `table.branches.checkout(name, version=v)` | `await branches.checkout(name, v)` |
Write column-level descriptions, tags, and logical groupings onto a LanceDB table's schema. Use this when the user wants to document, annotate, tag, or classify what their table columns ARE (embeddings vs labels vs eval metrics, model provenance, version families, etc.).
Works on local/OSS and remote Enterprise/Cloud tables alike — read the schema through the table handle, write through `update_field_metadata` (Python) / `updateFieldMetadata` (TypeScript).
## Metadata key conventions
All metadata uses namespaced keys:
| Key | Purpose | Example value |
|-----|---------|---------------|
| `lancedb:description` | Human-readable explanation of what the column contains | `"CLIP ViT-L/14 image embedding, L2-normalized (768-dim)"` |
| `lancedb:tag:<name>` | Flexible key-value tag; the suffix names the tag category | `lancedb:tag:field_type: "embedding"`, `lancedb:tag:model: "clip"`, `lancedb:tag:project_id: "foo"` |
| `lancedb:logical-column` | Logical group/family this column belongs to | `"clip_features"` |
Tags are open-ended — use whatever key suffix and value make sense given the user's intent. The tag suffix should describe *what is being classified* (e.g., `field_type`, `model`, `project_id`) and the value describes *how*. Multiple tags on the same column are fine — each is a separate key. All values are strings.
## Step 1: Read the schema and existing metadata
Read existing metadata before writing, to avoid redundant updates.
Python — `table.schema` (sync property; async: `await table.schema()`) returns a `pyarrow.Schema`. **Arrow field metadata is bytes-keyed in Python**:
```python
schema = table.schema
for field in schema:
meta = field.metadata or {} # dict[bytes, bytes], e.g. {b"lancedb:description": b"..."}
Quick method reference for Python LanceDB code. Cross-check source for non-trivial claims.
## Connect
If you're connecting to a remote database, use this:
```python
import lancedb
db = lancedb.connect("db://my-db", api_key=api_key, host_override=host_override) # remote
```
(values may be found in LANCEDB_API_KEY and LANCEDB_HOST_OVERRIDE, either in env vars or a .env file)
If you're connecting to a local table using OSS LanceDB, use this:
```python
db = lancedb.connect("./camelot-db") # local/OSS
```
If you're not sure which, or if you can't find the api_key or host_override params, ask the user.
**Place the local database directory next to the script/entrypoint that opens it** (i.e. resolve the path relative to the script, `Path(__file__).parent / "camelot-db"`), not buried under a shared `data/` folder. The Lance dataset is the database, not a data file — keeping it beside its code makes ownership obvious and paths stable regardless of the working directory the script is launched from.
**Do not name the directory `lancedb`** (e.g. `./lancedb`, `./data/lancedb`). It collides with the imported `lancedb` package name, which is confusing to read and easy to shadow in scripts. Give it a name derived from the repo or dataset with a clear prefix/suffix — for example `./<dataset>-db`, `./<repo>_lancedb`, or `./vectordb`.
| Full scan with filters/projection (sync) | `table.search().where(...).select(...).limit(...)` |
| Full scan with filters/projection (async) | `table.query().where(...).select(...).limit(...)` |
| Filter | `.where("col > 10")` |
| Projection | `.select(["id", "text"])` |
| Bound result count | `.limit(20)` |
| Collect bounded result as Python objects (default, no extra deps) | `.to_list()` on query/search result |
| Collect bounded result as Arrow (default, `pyarrow` always available) | `.to_arrow()` on query/search result |
| Collect bounded result as pandas (only if project uses pandas) | `.to_pandas()` on query/search result |
| Collect bounded result as Polars (only if project uses polars) | `.to_polars()` on query/search result |
## Sync vs Async Scan API
The plain-scan entry point differs between the sync and async clients. **Verified against `lancedb` 0.34.0** — re-check if the pinned version changes:
- **Sync** (`lancedb.connect(...)`): the table has **no `.query()` method**. Use `.search()` with no argument for a plain scan; it returns a query builder that supports `.where()`, `.select()`, `.limit()`, and the `.to_list()` / `.to_arrow()` / `.to_pandas()` / `.to_polars()` collectors.
Merges by default; a `None` value deletes that key; `"replace": True` swaps the whole map. Nested fields use dot-paths (`"a.b.c"`). `replace_field_metadata` is deprecated. See `references/column_metadata.md` for key conventions (`lancedb:description`, `lancedb:tag:<name>`, `lancedb:logical-column`) and the authoring workflow.
## Branches
```python
table.branches.list() # non-main branches; {} = only main
exp = table.branches.create("exp") # fork off main -> handle scoped to the branch
wip = db.open_table("t", branch="wip") # or open scoped directly
table.branches.delete("stale") # removes only the branch pointer
table.current_branch() # None = main
```
There is no global switch — scoping is per table handle: any read/write on a branch handle lands on that branch; the original handle keeps targeting main. See `references/branch_ops.md` for the model and isolation checks.
## Maintenance
```python
table.optimize()
```
Call this after every successful local/OSS ingestion. It handles compaction, cleanup of old versions according to retention, and index optimization. Do not add this for LanceDB Enterprise/Cloud remote tables; Enterprise handles compaction and cleanup automatically from cluster configuration.
Use these patterns when writing Python code with `lancedb`.
## Before Writing Code
Choose the output type from what the project actually depends on. **Do not assume `pandas` or `polars` is installed** — they are heavy dependencies that many LanceDB projects do not use. `pyarrow`, by contrast, ships as a LanceDB dependency and is always available, so it is a safe default to lean on.
Default output (after applying `select()` and `limit()`):
- **Python objects**: `.to_list()` — a list of dicts, no extra dependencies. Prefer this for scripts, examples, and agent-generated code unless there is a reason to do otherwise.
- **PyArrow**: `.to_arrow()` — a `pyarrow.Table`, when the surrounding code is Arrow-native or you need columnar/zero-copy handoff.
Only reach for a DataFrame when the project *already* declares that dependency:
- Pandas projects (pandas in `pyproject.toml`/requirements): `.to_pandas()`.
Use PyArrow schemas instead when the pipeline is already Arrow-native, needs
record-batch streaming, or has runtime schema requirements that would make a
Pydantic model harder to understand. Declare Pydantic as a direct project
dependency when application code imports it, even if LanceDB also depends on it.
## Recommended Patterns
### Bounded search or query
Use this for application reads, examples, notebooks, and agent-generated scripts:
```python
results = (
table.search(query_vector)
.where("status = 'ready'")
.select(["id", "text"])
.limit(20)
.to_list() # or .to_arrow(); .to_pandas()/.to_polars() only if the project uses them
)
```
Why: `search()` works across local and remote tables and on both the sync and async clients. `select()` avoids fetching unused columns. `limit()` prevents accidental full-table reads. `.to_list()` and `.to_arrow()` avoid assuming pandas/polars is installed (see "Before Writing Code").
For a **plain scan** (no query vector), the entry point differs by client:
```python
# Sync client: no .query() method — use .search() with no argument.
`table.query()` on a sync table raises `AttributeError` (verified on `lancedb` 0.34.0). See the "Sync vs Async Scan API" section in `api_reference.md`.
### Bounded query result conversion
It is fine to collect bounded query/search results:
df = table.search(query_vector).limit(10).to_pandas() # only if pandas is a project dep
```
### Local-only Lance dataset API
`table.to_lance()` does not itself materialize the full dataset. It returns the underlying `lance.LanceDataset`, making the table accessible through the PyLance dataset API. Use it when the task is explicitly local/OSS and needs Lance dataset methods not exposed by LanceDB:
```python
# Local/OSS only: RemoteTable does not expose table.to_lance().
ds = table.to_lance()
for batch in ds.to_batches(columns=["id", "text"], batch_size=10_000):
process(batch)
```
### Async Python
Keep the same shape and bound the result before collecting:
```python
results = await (
async_table.query()
.where("status = 'ready'")
.select(["id", "text"])
.limit(20)
.to_list() # or .to_arrow()
)
```
## Anti-Patterns
**Avoid the following anti-patterns in your code.**
### Table-level full materialization
Avoid whole-table collectors in portable or large-table code:
```python
df = table.to_pandas()
arrow_table = table.to_arrow()
polars_df = table.to_polars()
```
Why: local tables expose these whole-table collectors, but remote tables intentionally do not — a remote production table can be far larger than a local development table, so it is easy to accidentally pull the entire table into memory.
`table.to_lance()` is different: it is not a full materialization call, but it is still local/OSS-only and should not appear in code meant to run against remote Enterprise tables.
### Unbounded result collection
Avoid query/search collection without a meaningful limit:
Prefer `select(...).limit(...)` before collecting; for large reads, stream in batches instead.
### Per-row writes
Avoid loops that write one row per call:
```python
for row in rows:
table.add([row]) # one commit + fragment per row
```
Each `add()` creates a new version and fragment. Pass the whole batch in a single call, or chunk very large inputs:
```python
table.add(rows) # single commit
# for very large inputs, add batches of several thousand rows
```
After the final successful write to an embedded OSS table, call
`table.optimize()`. Skip this for Enterprise/Cloud tables because their
maintenance is automatic.
### Drop-then-reuse the same table name (Enterprise/Cloud)
Avoid dropping or overwriting a remote table and then reusing that name right away:
```python
db.drop_table("my_table")
table = db.create_table("my_table", data=rows) # reads 500 for ~5 min
table = db.create_table("my_table", data=rows, mode="overwrite") # same problem
```
Why: Enterprise/Cloud splits DDL (control plane) from query serving (data plane). The data plane caches the dataset behind a table name for up to `table_cache_ttl` (default 300s / 5 min), so after a drop/overwrite the DDL succeeds but queries against the reused name return `500 Internal Server Error` until the cache expires — and a fresh `describe` may still show the old schema. Instead, write to a **fresh name**, use `list_tables()` and fail if it already exists, then `rename_table(fresh, final)` onto the final name only after the old table's drop has propagated (~5 min). See the "Enterprise: never drop-then-reuse the same table name" section in `SKILL.md`. Local/OSS tables have no separate data plane — overwrite freely there.
### Guessing performance fixes
Avoid changing `nprobes`, `refine_factor`, or index types before checking the query plan and index stats. Diagnose first, then tune one knob at a time.
### Recommended: bulk ingestion for materialized data
```python
table.add(arrow_table)
table.add(df)
table.add(pa.dataset("data/", format="parquet"))
```
For very large initial loads, create the table empty first, then call `add(...)`. Passing data directly to `create_table(name, data)` can skip the auto-parallel write path.
### Recommended: iterator ingestion for generated or streamed data
Use chunks of several thousand rows or more when practical. Tiny batches and per-row writes create many small fragments.
### Anti-pattern: per-row `add()`
```python
for row in rows:
table.add([row])
```
Each call creates a version and fragment. This slows ingestion and later queries.
## Indexing
- Build a vector index once brute-force vector search becomes too slow. As a rule of thumb, local brute force is fine below roughly 100K vectors; beyond that, build an index.
- Use `IVF_PQ` as the general-purpose default. Enterprise builds this automatically.
- Use scalar indexes for filtered columns and merge/upsert keys.
- Use `BTREE` for mostly distinct numeric/string/temporal columns, `BITMAP` for booleans and low-cardinality columns, and `LABEL_LIST` for list membership queries.
- Keep full-text defaults unless phrase queries require position data.
Every remote (`db://...`) connection talks to such a server, and some operations
exist only there. In particular, all operations around jobs (listing, inspecting,
creating, or canceling jobs) run server-side — there is no local/OSS equivalent, so
resolve a server connection before attempting any job work. The job REST methods
themselves are documented in `references/remote_jobs.md`.
Every request needs two things:
1. **Base URL** — the server endpoint
2. **Credentials** — an API key (`x-api-key` header over REST), and usually a database name (`x-lancedb-database` header)
## Resolution steps
1. If the user already gave a URL and API key (or said which environment they're working against), use that.
2. Otherwise, look for credentials already available in the environment:
- Env vars like `LANCEDB_URI` / `LANCEDB_HOST` / `LANCEDB_API_KEY`
- A server endpoint already running or port-forwarded locally (the REST default port is 2333, i.e. `http://localhost:2333`)
3. If you didn't find both pieces, ask the user directly: **"What's your LanceDB endpoint's URL, and what's your API key?"** Also ask which database to use if it isn't obvious. Don't guess or probe further — the user knows their deployment.
## Validating the connection
Make a cheap authenticated request and check the status before starting real work:
Quick method reference for TypeScript LanceDB code. Cross-check source for non-trivial claims.
## Connect
```typescript
import * as lancedb from "@lancedb/lancedb";
const db = await lancedb.connect("./camelot-db");
```
**Place the local database directory next to the script/entrypoint that opens it** (resolve the path relative to the module, e.g. via `import.meta.dirname` / `__dirname`), not buried under a shared `data/` folder. The Lance dataset is the database, not a data file — keeping it beside its code makes ownership obvious and paths stable regardless of the working directory the script is launched from.
**Do not name the directory `lancedb`** (e.g. `./lancedb`, `./data/lancedb`). It collides with the imported `lancedb` package/namespace, which is confusing to read. Give it a name derived from the repo or dataset with a clear prefix/suffix — for example `./<dataset>-db`, `./<repo>_lancedb`, or `./vectordb`.
Remote connections use `db://...` plus Enterprise/Cloud credentials and deployment settings. Check current source/docs for exact connection options.
| Full scan with filters/projection | `table.query().where(...).select(...).limit(...)` |
| Filter | `.where("col > 10")` |
| Projection | `.select(["id", "text"])` |
| Bound result count | `.limit(20)` |
| Collect bounded result as objects | `.toArray()` on query/search result |
| Collect bounded result as Arrow | `.toArrow()` on query/search result |
| Stream result batches | `for await (const batch of table.query()...)` |
## Local vs Remote Safety
| API | Agent guidance |
| --- | --- |
| `table.search(...)` | Preferred read path |
| `table.query()` | Preferred scan/filter path |
| `await table.toArrow()` | Avoid in portable or large-table code |
| `await table.query().toArray()` with no `limit()` | Avoid; unbounded collection |
| `await table.query().toArrow()` with no `limit()` | Avoid; unbounded collection |
## Indexes
```typescript
await table.createIndex("vector");
await table.createIndex("status");
```
Use vector indexes for large vector search workloads and scalar indexes for filtered columns or merge/upsert keys. Check source/docs before specifying advanced index options.
Merges by default; a `null` value deletes that key; `replace: true` swaps the whole map. Nested fields use dot-paths (`"a.b.c"`). See `references/column_metadata.md` for key conventions (`lancedb:description`, `lancedb:tag:<name>`, `lancedb:logical-column`) and the authoring workflow.
const wip2 = await db.openTable("t", { branch: "wip" }); // or open scoped directly
await branches.delete("stale"); // removes only the branch pointer
table.currentBranch(); // null = main
```
There is no global switch — scoping is per table handle: any read/write on a branch handle lands on that branch; the original handle keeps targeting main. See `references/branch_ops.md` for the model and isolation checks.
## Maintenance
```typescript
await table.optimize();
```
Call this after every successful local/OSS ingestion. It handles compaction, cleanup of old versions according to retention, and index optimization. Do not add this for LanceDB Enterprise/Cloud remote tables; Enterprise handles compaction and cleanup automatically from cluster configuration.
Use these patterns when writing TypeScript code with `@lancedb/lancedb`.
## Recommended Patterns
### Bounded query
Use this for application reads, scripts, and examples:
```typescript
const rows = await table
.query()
.where("status = 'ready'")
.select(["id", "text"])
.limit(20)
.toArray();
```
### Bounded vector search
```typescript
const rows = await table
.search(queryVector)
.select(["id", "text"])
.limit(20)
.toArray();
```
### Batch streaming for larger reads
When the task needs many rows, avoid collecting everything at once:
```typescript
for await (const batch of table
.query()
.where("status = 'ready'")
.select(["id", "text"])
.limit(10_000)) {
process(batch);
}
```
## Anti-Patterns
**Avoid the following anti-patterns in your code.**
### Table-level full materialization
Avoid whole-table collectors in portable or large-table code:
```typescript
const tableArrow = await table.toArrow();
```
Why: local tables expose these whole-table collectors, but remote tables intentionally do not — a remote production table can be far larger than a local development table, so it is easy to accidentally pull the entire table into memory.
### Unbounded result collection
Avoid query/search collection without a meaningful limit:
Prefer `select(...).limit(...)` before collecting; for large reads, stream in batches instead.
### Per-row writes
Avoid loops that write one row per call:
```typescript
for (const row of rows) {
await table.add([row]); // one commit + fragment per row
}
```
Each `add()` creates a new version and fragment. Pass the whole batch in a single call, or chunk very large inputs:
```typescript
await table.add(rows); // single commit
// for very large inputs, add in chunks of several thousand rows
```
### Drop-then-reuse the same table name (Enterprise/Cloud)
Avoid dropping or overwriting a remote table and then reusing that name right away:
```typescript
await db.dropTable("my_table");
const table = await db.createTable("my_table", rows); // reads 500 for ~5 min
const table = await db.createTable("my_table", rows, { mode: "overwrite" }); // same problem
```
Why: Enterprise/Cloud splits DDL (control plane) from query serving (data plane). The data plane caches the dataset behind a table name for up to `table_cache_ttl` (default 300s / 5 min), so after a drop/overwrite the DDL succeeds but queries against the reused name return `500 Internal Server Error` until the cache expires — and a fresh `describe` may still show the old schema. Instead, write to a **fresh name**, use `tableNames()` and fail if it already exists, then `renameTable(fresh, final)` onto the final name only after the old table's drop has propagated (~5 min). See the "Enterprise: never drop-then-reuse the same table name" section in `SKILL.md`. Local/OSS tables have no separate data plane — overwrite freely there.
### Guessing performance fixes
Avoid changing `nprobes`, `refineFactor`, `ef`, or index settings before checking `analyzePlan()` and `indexStats(...)`. Diagnose first, then tune one knob at a time.
Use this when writing TypeScript code that ingests data, queries large tables, builds indexes, or investigates latency.
## Ingestion
- Prefer bulk or batched writes.
- Avoid per-row write loops; they create many small commits/fragments.
- For generated data, accumulate reasonable batches before adding.
- For file-backed data, prefer APIs that stream from Arrow/Parquet-style inputs when available.
## Indexing
- Build a vector index once brute-force vector search becomes too slow. As a rule of thumb, local brute force is fine below roughly 100K vectors; beyond that, build an index.
- Use the general-purpose vector index defaults unless the task has explicit recall/latency requirements.
- Build scalar indexes for filtered columns and merge/upsert keys.
- Use full-text index phrase options only when phrase queries require them.
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Will Jones
Yang Cen
Lance Release
buduoqiu
kid
dependabot[bot]
Dan Tasse
Xuanwo
Vivek
LanceDB Robot
Heng Ge
Andrew Chen
Eran Dagan
Prashanth Rao
Expyron
Mateusz Szewczyk
Drew Gallardo
Vitaliy
Kobi Hikri
Tobias
Jack Ye
Weston Pace
Mark McDonald
