This wires Lance's existing `jieba/*` and `lindera/*` native FTS
tokenizers through the Python SDK instead of leaving them behind
disabled features and narrow public typing. It also documents the
`LANCE_LANGUAGE_MODEL_HOME` model layout and adds Python coverage for
successful CJK indexing plus missing-model error guidance.
Closes#2168.
## Summary
PyTorch's `DataLoader` uses fork-based multiprocessing by default on
Linux, but threads do not survive `fork()`. LanceDB's Python bindings
drive async work through two threaded layers, both of which become inert
in a forked child:
- `BackgroundEventLoop` runs an asyncio loop on a Python
`threading.Thread`.
- `pyo3-async-runtimes::tokio` holds a global multi-threaded tokio
runtime whose worker threads also die on fork — and its runtime lives in
a `OnceLock` that cannot be replaced after first use.
As a result, any `Permutation` (or other async API) used inside a
fork-based `DataLoader` worker hangs indefinitely. This PR makes both
layers fork-safe so `Permutation` works as a `torch.utils.data.Dataset`
with `num_workers > 0`.
## Approach
### Rust — new `python/src/runtime.rs`
Mirrors the pattern used in [Lance's Python
bindings](456198cd6f/python/src/lib.rs (L139)),
adapted for the async-bridge use case.
- `LanceRuntime` implements `pyo3_async_runtimes::generic::Runtime +
ContextExt`, backed by an `AtomicPtr<tokio::runtime::Runtime>` we own
(sidestepping `pyo3-async-runtimes`'s frozen `OnceLock` global).
- A `pthread_atfork(after_in_child)` handler nulls the pointer; the next
`spawn` rebuilds the runtime in the child. The previous runtime is
intentionally **leaked** — calling `Drop` would try to join now-dead
worker threads and hang.
- `runtime::future_into_py` is a drop-in for
`pyo3_async_runtimes::tokio::future_into_py`. All ~80 call sites in
`arrow.rs` / `connection.rs` / `permutation.rs` / `query.rs` /
`table.rs` are updated to route through it.
- `python/Cargo.toml` adds `libc = "0.2"` and the tokio
`rt-multi-thread` feature.
### Python — `lancedb/background_loop.py`
- Refactors `BackgroundEventLoop.__init__` to a reusable `_start()`
method.
- An `os.register_at_fork(after_in_child=…)` hook calls `LOOP._start()`
to give the singleton a fresh asyncio loop and thread **in place**. This
matters because the rest of the codebase imports `LOOP` via `from
.background_loop import LOOP` — rebinding the module attribute would
leave those references holding the dead loop.
### Python — `lancedb/__init__.py`
Removes the `__warn_on_fork` pre-fork warning (and the now-unused
`import warnings`). Fork is supported.
## Test plan
- [x] New `test_permutation_dataloader_fork_workers` in
`python/tests/test_torch.py`: runs a `Permutation` through
`torch.utils.data.DataLoader(num_workers=2,
multiprocessing_context="fork")` inside a spawn-isolated child with a
30s hang detector. **Pre-fix**: timed out at 36s. **Post-fix**: passes
in ~3.6s.
- [x] New `test_remote_connection_after_fork` in
`python/tests/test_remote_db.py`: forks a child that creates a fresh
`lancedb.connect(...)` against a mock HTTP server and calls
`table_names()`; passes in <1s, validates the runtime reset is
sufficient for fresh remote clients.
- [x] All 62 tests in `test_torch.py` + `test_permutation.py` pass.
- [x] All 35 tests in `test_remote_db.py` pass.
- [x] `test_table.py` (87) + `test_db.py` + `test_query.py` (157, minus
one unrelated `sentence_transformers` import skip) — 244 passing.
- [x] `cargo clippy -p lancedb-python --tests` clean.
- [x] `cargo fmt`, `ruff check`, `ruff format` all clean.
## Known limitation (follow-up)
This PR makes a **freshly-built** `lancedb.connect(...)` work in a
forked child. An **inherited** `Connection` from the parent still
carries an inherited `reqwest::Client` whose hyper connection pool
references socket FDs and TCP/TLS state shared with the parent — using
it from the child after fork is unsafe (especially with HTTP/1.1
keep-alive). The recommended pattern for fork-based `DataLoader` workers
that hit a remote DB is to construct a new connection inside the worker.
Auto-clearing inherited HTTP client pools on fork would require tracking
live `Connection` instances in `lancedb` core and is left for a
follow-up PR.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
When pytorch is used with multiprocessing and the mp mode is spawn then
the Permutation needs to be pickled. It could not be pickled because
`Table` and `Connection` are not serializable. This PR adds pickle
support to Permutation without adding general pickle support to `Table`
or `Connection`. To add general support we probably need to start by
adding serialization in the namespace client.
In the meantime this PR enable pickling by adding special cases for:
* In-memory tables (just serialize as Arrow IPC)
* Native tables (serialize the URI)
If a user is not using one of the above cases (e.g. using a remote
connection) then they will need to provide a connection factory that can
be pickled.
## Breaking change
`PermutationBuilder.persist(...)` is removed from the Python bindings;
the permutation table is now always in-memory. The underlying Rust
`PermutationBuilder::persist` API is untouched and can be re-exposed
later if needed. It probably won't make sense to do that until we have a
way to serialize `Table` and `Connection`.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hi, the hybrid query error message looks like it can use a space, just
added it.
```python
def _validate_query(self, query, vector=None, text=None):
if query is not None and (vector is not None or text is not None):
raise ValueError(
"You can either provide a string query in search() method"
"or set `vector()` and `text()` explicitly for hybrid search."
"But not both."
)
```
Adds manifest_enabled for local/native connections so directory
namespace manifests can be the source of truth, including migration from
directory listing and Azure credential vending feature wiring. Also
exposes the option through Rust, Python, and Node bindings with focused
validation.
This follows the Rust-side Tantivy removal by deleting the remaining
Python Tantivy runtime, tests, and packaging references.
It also turns the legacy Python-only Tantivy parameters into explicit
errors and stops reading legacy `_indices/fts` directories so Python FTS
is fully native-only.
So far, I have been using a hacky approach that creates and opens
namespace-backed table, by getting its location and use a temporary
lancedb connection to create or open it. This was working for features
like credentials vending but is no longer fully working for the managed
versioning feature, recently geneva tests have been failing here and
there and various patches are not addressing the root cause. This PR
fully fixes this and implements proper rust binding for it.
Specifically:
- build a real Rust namespace-backed connection from the Python
namespace client
- route namespace table create/open through that connection instead of
resolved-location temp connections
- keep namespace client naming consistent in the Rust bridge and
preserve federated namespace + DuckDB behavior
## Summary
- pass `namespace_client` through the Python create-table path
- ensure schema-only namespace table creation uses the namespace-aware
empty-table flow
- fix reopening namespace tables created without initial data
## Summary
- delegate child-namespace `ListingDatabase` operations through an
eagerly initialized `LanceNamespaceDatabase`
- support nested namespace create/open/list/drop flows without requiring
callers to inject explicit locations
- add `namespace_client_properties` plumbing for local and namespace
connections so directory namespace settings like
`table_version_tracking_enabled` can be configured
- add regression tests for nested namespace ops and namespace client
property propagation
## Summary
Add connection serialization and child namespace support to
`LanceDBConnection`.
- `DBConnection.serialize()` / `lancedb.deserialize()` for connection
reconstruction in remote workers
- Cache `namespace_client()` in `LanceDBConnection` to avoid repeated
DirectoryNamespace builds
- `LanceDBConnection` transparently delegates child namespace operations
(open_table, create_table, list_tables, drop_table, create_namespace,
etc.) to `LanceNamespaceDBConnection` via `_namespace_conn()`
- Root namespace operations still go through the original Rust path
- Generic worker property override mechanism: any
`namespace_client_properties` key prefixed with `_lancedb_worker_` has
the prefix stripped and overrides the corresponding property when
`deserialize(data, for_worker=True)`
- `LanceNamespaceDBConnection` stores
`namespace_client_impl`/`namespace_client_properties` for serialization
roundtrip
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
## Summary
- migrate gemini-text embedding provider from deprecated
google.generativeai to google.genai
- update Python embedding extra dependency to google-genai
- update default model name to gemini-embedding-001
- adapt embed calls to Client().models.embed_content(...)
- apply lint fixes from CI
## Related
- Closes#3191
`.get(b"split_names", None).decode()` was called unconditionally in both
Permutations.__init__ and Permutation.from_tables(), crashing with
AttributeError when schema metadata existed but lacked the split_names
key. Guard the decode behind a None check and add regression tests.
## Problem
`on_bad_vectors="drop"` is supposed to remove invalid vector rows before
write, but for some schema-defined vector columns it can still fail
later during Arrow cast instead of dropping the bad row.
Repro:
```python
class MySchema(LanceModel):
text: str
embedding: Vector(16)
table = db.create_table("test", schema=MySchema)
table.add(
[
{"text": "hello", "embedding": []},
{"text": "bar", "embedding": [0.1] * 16},
],
on_bad_vectors="drop",
)
```
Before:
```
RuntimeError
Arrow error: C Data interface error: Invalid: ListType can only be casted to FixedSizeListType if the lists are all the expected size.
```
After:
```
rows 1
texts ['bar']
```
## Solution
Make bad-vector sanitization use schema dimensions before cast, while
keeping the handling scoped to vector columns identified by schema
metadata or existing vector-name heuristics.
This also preserves existing integer vector inputs and avoids applying
on_bad_vectors to unrelated fixed-size float columns.
Fixes#1670
Signed-off-by: yaommen <myanstu@163.com>
## Summary
- Add a `user_id` field to `ClientConfig` that allows users to identify
themselves to LanceDB Cloud/Enterprise
- The user_id is sent as the `x-lancedb-user-id` HTTP header in all
requests
- Supports three configuration methods:
- Direct assignment via `ClientConfig.user_id`
- Environment variable `LANCEDB_USER_ID`
- Indirect env var lookup via `LANCEDB_USER_ID_ENV_KEY`
Closes#3230🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
## Summary
Fixes#1846.
Python `Enum` fields raised `TypeError: Converting Pydantic type to
Arrow Type: unsupported type <enum 'SomethingTypes'>` when converting a
Pydantic model to an Arrow schema.
The fix adds Enum detection in `_pydantic_type_to_arrow_type`. When an
Enum subclass is encountered, the value type of its members is inspected
and mapped to the appropriate Arrow type:
- `str`-valued enums (e.g. `class Status(str, Enum)`) → `pa.utf8()`
- `int`-valued enums (e.g. `class Priority(int, Enum)`) → `pa.int64()`
- Other homogeneous value types → the Arrow type for that Python type
- Mixed-value or empty enums → `pa.utf8()` (safe fallback)
This covers the common `(str, Enum)` and `(int, Enum)` mixin patterns
used in practice.
## Changes
- `python/python/lancedb/pydantic.py`: add Enum branch in
`_pydantic_type_to_arrow_type`
- `python/python/tests/test_pydantic.py`: add `test_enum_types` covering
`str`, `int`, and `Optional` Enum fields
## Note on #2395
PR #2395 handles `StrEnum` (Python 3.11+) specifically, using a
dictionary-encoded type. This PR handles the broader `(str, Enum)` /
`(int, Enum)` mixin pattern that works across all Python versions and
stores values as their natural Arrow type.
AI assistance was used in developing this fix.
1. Refactored every client (Rust core, Python, Node/TypeScript) so
“namespace” usage is explicit: code now keeps namespace paths
(namespace_path) separate from namespace clients (namespace_client).
Connections propagate the client, table creation routes through it, and
managed versioning defaults are resolved from namespace metadata. Python
gained LanceNamespaceDBConnection/async counterparts, and the
namespace-focused tests were rewritten to match the clarified API
surface.
2. Synchronized the workspace with Lance 5.0.0-beta.3 (see
https://github.com/lance-format/lance/pull/6186 for the upstream
namespace refactor), updating Cargo/uv lockfiles and ensuring all
bindings align with the new namespace semantics.
3. Added a namespace-backed code path to lancedb.connect() via new
keyword arguments (namespace_client_impl, namespace_client_properties,
plus the existing pushdown-ops flag). When those kwargs are supplied,
connect() delegates to connect_namespace, so users can opt into
namespace clients without changing APIs. (The async helper will gain
parity in a later change)
The test added in #3190 unconditionally imports `PIL`, which is an
optional dependency. This causes CI failures in environments where
Pillow isn't installed (`ModuleNotFoundError: No module named 'PIL'`).
Use `pytest.importorskip` to skip gracefully when Pillow is unavailable.
Fixes CI failure on main.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
## Summary
- Namespace tests expected `RuntimeError` for table-not-found and
namespace-not-empty cases, but `lance_namespace` raises
`TableNotFoundError` and `NamespaceNotEmptyError` which inherit from
`Exception`, not `RuntimeError`.
- Updated `pytest.raises` to use the correct exception types.
## Test plan
- [x] CI passes on `test_namespace.py`
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
url_retrieve() calls urllib.request.urlopen() but only urllib.error was
imported, causing AttributeError for any HTTP URL input. This affects
open-clip, siglip, and jinaai embedding functions when processing image
URLs.
The bug has existed since the embeddings API refactor (#580) but was
masked because most users pass local file paths or bytes rather than
HTTP URLs.
Fixes#3183
## Summary
When `table.add(mode='overwrite')` is called, PyArrow infers input data
types (e.g. `list<double>`) which differ from the original table schema
(e.g. `fixed_size_list<float32>`). Previously, overwrite mode bypassed
`cast_to_table_schema()` entirely, so the inferred types replaced the
original schema, breaking vector search.
This fix builds a merged target schema for overwrite: columns present in
the existing table schema keep their original types, while columns
unique to the input pass through as-is. This way
`cast_to_table_schema()` is applied unconditionally, preserving vector
column types without blocking schema evolution.
## Changes
- `rust/lancedb/src/table/add_data.rs`: For overwrite mode, construct a
target schema by matching input columns against the existing table
schema, then cast. Non-overwrite (append) path is unchanged.
- Added `test_add_overwrite_preserves_vector_type` test that creates a
table with `fixed_size_list<float32>`, overwrites with `list<double>`
input, and asserts the original type is preserved.
## Test Plan
- `cargo test --features remote -p lancedb -- test_add_overwrite` — all
4 overwrite tests pass
- Full suite: 454 passed, 2 failed (pre-existing `remote::retry` flakes
unrelated to this change)
---------
Signed-off-by: majiayu000 <1835304752@qq.com>
dict.update() mutates in place and returns None. Assigning its result
caused with_metadata(None) to strip all schema metadata when embedding
metadata was merged during create_table with embedding_functions.
## Summary
Adds progress reporting for `table.add()` so users can track large write
operations. The progress callback is available in Rust, Python (sync and
async), and through the PyO3 bindings.
### Usage
Pass `progress=True` to get an automatic tqdm bar:
```python
table.add(data, progress=True)
# 100%|██████████| 1000000/1000000 [00:12<00:00, 82345 rows/s, 45.2 MB/s | 4/4 workers]
```
Or pass a tqdm bar for more control:
```python
from tqdm import tqdm
with tqdm(unit=" rows") as pbar:
table.add(data, progress=pbar)
```
Or use a callback for custom progress handling:
```python
def on_progress(p):
print(f"{p['output_rows']}/{p['total_rows']} rows, "
f"{p['active_tasks']}/{p['total_tasks']} workers, "
f"done={p['done']}")
table.add(data, progress=on_progress)
```
In Rust:
```rust
table.add(data)
.progress(|p| println!("{}/{:?} rows", p.output_rows(), p.total_rows()))
.execute()
.await?;
```
### Details
- `WriteProgress` struct in Rust with getters for `elapsed`,
`output_rows`, `output_bytes`, `total_rows`, `active_tasks`,
`total_tasks`, and `done`. Fields are private behind getters so new
fields can be added without breaking changes.
- `WriteProgressTracker` tracks progress across parallel write tasks
using a mutex for row/byte counts and atomics for active task counts.
- Active task tracking uses an RAII guard pattern (`ActiveTaskGuard`)
that increments on creation and decrements on drop.
- For remote writes, `output_bytes` reflects IPC wire bytes rather than
in-memory Arrow size. For local writes it uses in-memory Arrow size as a
proxy (see TODO below).
- tqdm postfix displays throughput (MB/s) and worker utilization
(active/total).
- The `done` callback always fires, even on error (via `FinishOnDrop`),
so progress bars are always finalized.
### TODO
- Track actual bytes written to disk for local tables. This requires
Lance to expose a progress callback from its write path. See
lance-format/lance#6247.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
## Problem
The generated Python API docs for
`lancedb.table.IndexStatistics.index_type` were misleading because
mkdocstrings renders that field’s type annotation directly, and the
existing `Literal[...]` listed only a subset of the actual canonical SDK
index type strings.
Current (missing index types):
<img width="823" height="83" alt="image"
src="https://github.com/user-attachments/assets/f6f29fe3-4c16-4d00-a4e9-28a7cd6e19ec"
/>
## Fix
- Update the `IndexStatistics.index_type` annotation in
`python/python/lancedb/table.py` to include the full supported set of
canonical values, so the generated docs show all valid index_type
strings inline.
- Add a small regression test in `python/python/tests/test_index.py` to
ensure the docs-facing annotation does not drift silently again in case
we add a new index/quantization type in the future.
- Bumps mkdocs and material theme versions to mkdocs 1.6 to allow access
to more features like hooks
After fix (all index types are included and tested for in the
annotations):
<img width="1017" height="93" alt="image"
src="https://github.com/user-attachments/assets/66c74d5c-34b3-4b44-8173-3ee23e3648ac"
/>
When using hybrid search with a where filter, the prefilter argument is
silently inverted. Passing prefilter=True actually performs
post-filtering, and prefilter=False actually performs pre-filtering.
## Summary
- Removes the "Experimental API" section from `optimize` method
documentation across Rust, Python, and TypeScript
- Adds a warning to `delete_unverified` documentation in all bindings:
this should only be set to true if you can guarantee no other process is
working on the dataset, otherwise it could be corrupted
- Fixes a typo ("shoudl" → "should")
Closes#3125🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Implement `RemoteTable.prewarm_data(columns)` calling `POST
/v1/table/{id}/page_cache/prewarm/`
- Implement `RemoteTable.prewarm_index(name)` calling `POST
/v1/table/{id}/index/{name}/prewarm/` (previously returned
`NotSupported`)
- Add `BaseTable::prewarm_data(columns)` trait method and `Table` public
API in Rust core
- Add PyO3 bindings and Python API (`AsyncTable`, `LanceTable`,
`RemoteTable`) for `prewarm_data`
- Add type stubs for `prewarm_index` and `prewarm_data` in
`_lancedb.pyi`
- Upgrade Lance to 3.0.0-rc.3 with breaking change fixes
Co-authored-by: Will Jones <willjones127@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Without this fix, if user directly use the native table to do operations
like `add_columns`, even if it is configured to use namespace db
connection, it is not really propagated through.
The fix is to bring lancedb's python binding up to date and do a similar
implementation as https://github.com/lance-format/lance/pull/5968, and
make sure the namespace is fully propagated through all the related
calls.
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
When we write data with `add()`, we can input data to the table's
schema. However, we were using "safe" mode, which propagates errors as
nulls. For example, if you pass `u64::max` into a field that is a `u32`,
it will just write null instead of giving overflow error. Now it
propagates the overflow. This is the same behavior as other systems like
DuckDB.
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
## Summary
- Add `@value_to_sql.register(dict)` handler that converts Python dicts
to DataFusion's `named_struct()` SQL syntax
- Enables updating struct-typed columns via `table.update(values={"col":
{"field_a": 1, "field_b": "hello"}})`
- Recursively handles nested structs, lists, nulls, and all existing
scalar types
Closes#1363
## Details
The `named_struct` function was introduced in DataFusion 38 and is now
available (LanceDB uses DataFusion 52.1). The implementation follows the
existing `singledispatch` pattern in `util.py`.
**Example conversion:**
```python
value_to_sql({"field_a": 1, "field_b": "hello"})
# => "named_struct('field_a', 1, 'field_b', 'hello')"
```
## Test plan
- [x] Unit tests for flat struct, nested struct, list inside struct,
mixed types, null values, and empty dict
- [ ] CI integration tests with actual table.update() on struct columns
🔗 [DataFusion named_struct
docs](https://datafusion.apache.org/user-guide/sql/scalar_functions.html#named-struct)
We don't necessarily need to do this, but one user was confused having
used `fast_search=True` as a keyword argument for vector searches, but
being unable to do so for FTS, even after the most recent changes. I
think this is the only discrepancy in where that is possible.
This hooks up a new writer implementation for the `add()` method. The
main immediate benefit is it allows streaming requests to remote tables,
and at the same time allowing retries for most inputs.
In NodeJS, we always convert the data to `Vec<RecordBatch>`, so it's
always retry-able.
For Python, all are retry-able, except `Iterator` and
`pa.RecordBatchReader`, which can only be consumed once. Some, like
`pa.datasets.Dataset` are retry-able *and* streaming.
A lot of the changes here are to make the new DataFusion write pipeline
maintain the same behavior as the existing Python-based preprocessing,
such as:
* casting input data to target schema
* rejecting NaN values if `on_bad_vectors="error"`
* applying embedding functions.
In future PRs, we'll enhance these by moving the embedding calls into
DataFusion and making sure we parallelize them. See:
https://github.com/lancedb/lancedb/issues/3048
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
Fixes#1679
This PR prevents the OpenAI embedding function from retrying when
receiving a 401 Unauthorized error. Authentication errors are permanent
failures that won't be fixed by retrying, yet the current implementation
retries all exceptions up to 7 times by default.
## Changes
- Modified `retry_with_exponential_backoff` in `utils.py` to check for
non-retryable errors before retrying
- Added `_is_non_retryable_error` helper function that detects:
- Exceptions with name `AuthenticationError` (OpenAI's 401 error)
- Exceptions with `status_code` attribute of 401 or 403
- Enhanced OpenAI embeddings to explicitly catch and re-raise
`AuthenticationError` with better logging
- Added unit test `test_openai_no_retry_on_401` to verify authentication
errors don't trigger retries
## Test Plan
- Added test that verifies:
1. A function raising `AuthenticationError` is only called once
2. No retry delays occur (sleep is never called)
- Existing tests continue to pass
- Formatting applied via `make format`
## Example Behavior
**Before**: With an invalid API key, users would see 7 retry attempts
over ~2 minutes:
```
WARNING:root:Error occurred: Error code: 401 - {'error': {'message': 'Incorrect API key provided...'}}
Retrying in 3.97 seconds (retry 1 of 7)
WARNING:root:Error occurred: Error code: 401...
Retrying in 7.94 seconds (retry 2 of 7)
...
```
**After**: With an invalid API key, the error is raised immediately:
```
ERROR:root:Authentication failed: Invalid API key provided
AuthenticationError: Error code: 401 - {'error': {'message': 'Incorrect API key provided...'}}
```
This provides better UX and prevents unnecessary API calls that would
fail anyway.
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
There are old and outdated files in our embedding registry that can
confuse coding agents. This PR deprecates the following files that have
newer, more modern methods to generate such embeddings.
- Deprecate `embeddings/siglip.py`
- Deprecate `embeddings/gte.py`
## Why this change?
Per a discussion with @AyushExel, the [embedding registry directory
](1840aa7edc/python/python/lancedb/embeddings)
in the LanceDB repo has a number of outdated files that need to be
deprecated.
See https://github.com/lancedb/docs/issues/85 for the docs gaps that
identified this.
- Add note in `openclip` docs that it can be used for SigLip embeddings,
which it now supports
- Add note in the `sentence-transformers` page that ALL text embedding
models on Hugging Face can be used
## Problem
When applying hard filters that result in zero matches, hybrid search
crashes with `IndexError: list index out of range` during reranking.
This happens because empty result tables are passed through the full
reranker pipeline, which expects at least one result.
Traceback from the issue:
```
lancedb/query.py: in _combine_hybrid_results
results = reranker.rerank_hybrid(fts_query, vector_results, fts_results)
lancedb/rerankers/answerdotai.py: in rerank_hybrid
combined_results = self._rerank(combined_results, query)
...
IndexError: list index out of range
```
## Fix
Added an early return in `_combine_hybrid_results` when both vector and
FTS results are empty. Instead of passing empty tables through
normalization, reranking, and score restoration (which can fail in
various ways), we now build a properly-typed empty result table with the
`_relevance_score` column and return it directly.
## Test
Added `test_empty_hybrid_result_reranker` that exercises
`_combine_hybrid_results` directly with empty vector and FTS tables,
verifying:
- Returns empty table with correct schema
- Includes `_relevance_score` column
- Respects `with_row_ids` flag
Closes#2425
This changes around the output format of `Permutation` in some breaking
ways but I think the API is still new enough to be considered
experimental.
1. In order to align with both huggingface's dataset and torch's
expectations the default output format is now a list of dicts
(row-major) instead of a dict of lists (column-major). I've added a
python_col option which will return the dict of lists.
2. In order to align with pytorch's expectation the `torch` format is
now a list of tensors (row-major) instead of a 2D tensor (column-major).
I've added a torch_col option which will return the 2D tensor instead.
Added tests for torch integration with Permutation
~~Leaving draft until https://github.com/lancedb/lancedb/pull/3013
merges as this is built on top of that~~
Closes#3000
The hybrid search `explain_plan` now shows the reranker as the top-level
node with
the vector and FTS sub-plans indented underneath, instead of just
listing them
separately with no reranker context.
**Before:**
```
Vector Search Plan:
ProjectionExec: ...
FTS Search Plan:
ProjectionExec: ...
```
**After:**
```
RRFReranker(K=60)
Vector Search Plan:
ProjectionExec: ...
FTS Search Plan:
ProjectionExec: ...
```
Other rerankers display similarly ; e.g.
`LinearCombinationReranker(weight=0.7, fill=1.0)`,
`MRRReranker(weight_vector=0.5, weight_fts=0.5)`,
`CohereReranker(model_name=name)`.
---------
Signed-off-by: dask-58 <googldhruv@gmail.com>
Co-authored-by: Will Jones <willjones127@gmail.com>
Fixes#2999
The error message previously said `"field_names must be a string when
use_tantivy=False"` implying they should use the to be deprecated
tantivy backend #2998.
Updated the error message and docstring to instead guide users to create
a separate FTS index for each field
Signed-off-by: dask-58 <googldhruv@gmail.com>
Expose `initial_storage_options()` and `latest_storage_options()` in
lance Dataset, in lancedb rust, python and typescript SDKs.
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
