This PR adds ColPali support with ColPaliEmbeddings class (tagged
"colpali") using ColQwen2.5 for multi-vector text/image embeddings. Also
added MultiVector Pydantic type to handle the vector lists.
I've added some integration test for the embedding model and some unit
test for the new Pydantic type. Could be a template for other ColPali
variants as well. or until transformers🤗 starts supporting it.
Still `TODO`:
- [ ] Documentation
- [ ] Add an example
_Could also allow Image as query, but didn't work well when testing it._
[ColPali-Engine](https://github.com/illuin-tech/colpali) version:
0.3.9.dev17+g3faee24
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **New Features**
- Introduced support for ColPali-based multimodal multi-vector
embeddings for both text and images.
- Added a new embedding class for generating multi-vector embeddings,
configurable for various model and processing options.
- Added a new Pydantic type for multi-vector embeddings, supporting
validation and schema generation for lists of fixed-dimension vectors.
- **Bug Fixes**
- Ensured proper asynchronous index creation in query tests for improved
reliability.
- **Tests**
- Added integration tests for ColPali embeddings, including
text-to-image search and validation of multi-vector fields.
- Added comprehensive tests for the new multi-vector Pydantic type,
covering schema, validation, and default value behavior.
- **Chores**
- Updated optional dependencies to include the ColPali engine.
- Added utility to check for availability of flash attention support.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
Closes#2287
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **New Features**
- Added configurable timeout support for query executions. Users can now
specify maximum wait times for queries, enhancing control over
long-running operations across various integrations.
- **Tests**
- Expanded test coverage to validate timeout behavior in both
synchronous and asynchronous query flows, ensuring timely error
responses when query execution exceeds the specified limit.
- Introduced a new test suite to verify query operations when a timeout
is reached, checking for appropriate error handling.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
add analyze plan api to allow executing the queries and see runtime
metrics.
Which help identify the query IO overhead and help identify query
slowness
This PR adds a `to_query_object` method to the various query builders
(except not hybrid queries yet). This makes it possible to inspect the
query that is built.
In addition this PR does some normalization between the sync and async
query paths. A few custom defaults were removed in favor of None (with
the default getting set once, in rust).
Also, the synchronous to_batches method will now actually stream results
Also, the remote API now defaults to prefiltering
@wjones127 is there a standard way you guys setup your virtualenv? I can
either relist all the dependencies in the pyright precommit section, or
specify a venv, or the user has to be in the virtual environment when
they run git commit. If the venv location was standardized or a python
manager like `uv` was used it would be easier to avoid duplicating the
pyright dependency list.
Per your suggestion, in `pyproject.toml` I added in all the passing
files to the `includes` section.
For ruff I upgraded the version and removed "TCH" which doesn't exist as
an option.
I added a `pyright_report.csv` which contains a list of all files sorted
by pyright errors ascending as a todo list to work on.
I fixed about 30 issues in `table.py` stemming from str's being passed
into methods that required a string within a set of string Literals by
extracting them into `types.py`
Can you verify in the rust bridge that the schema should be a property
and not a method here? If it's a method, then there's another place in
the code where `inner.schema` should be `inner.schema()`
``` python
class RecordBatchStream:
@property
def schema(self) -> pa.Schema: ...
```
Also unless the `_lancedb.pyi` file is wrong, then there is no
`__anext__` here for `__inner` when it's not an `AsyncGenerator` and
only `next` is defined:
``` python
async def __anext__(self) -> pa.RecordBatch:
return await self._inner.__anext__()
if isinstance(self._inner, AsyncGenerator):
batch = await self._inner.__anext__()
else:
batch = await self._inner.next()
if batch is None:
raise StopAsyncIteration
return batch
```
in the else statement, `_inner` is a `RecordBatchStream`
```python
class RecordBatchStream:
@property
def schema(self) -> pa.Schema: ...
async def next(self) -> Optional[pa.RecordBatch]: ...
```
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
Reviving #1966.
Closes#1938
The `search()` method can apply embeddings for the user. This simplifies
hybrid search, so instead of writing:
```python
vector_query = embeddings.compute_query_embeddings("flower moon")[0]
await (
async_tbl.query()
.nearest_to(vector_query)
.nearest_to_text("flower moon")
.to_pandas()
)
```
You can write:
```python
await (await async_tbl.search("flower moon", query_type="hybrid")).to_pandas()
```
Unfortunately, we had to do a double-await here because `search()` needs
to be async. This is because it often needs to do IO to retrieve and run
an embedding function.
we found a bug that flat KNN plan node's stats is not in right order as
fields in schema, it would cause an error if querying with distance
range and new unindexed rows.
we've fixed this in lance so add this test for verifying it works
Signed-off-by: BubbleCal <bubble-cal@outlook.com>
This PR aims to fix#2047 by doing the following things:
- Add a distance_type parameter to the sync query builders of Python
SDK.
- Make metric an alias to distance_type.
This includes several improvements and fixes to the Python Async query
builders:
1. The API reference docs show all the methods for each builder
2. The hybrid query builder now has all the same setter methods as the
vector search one, so you can now set things like `.distance_type()` on
a hybrid query.
3. Re-rankers are now properly hooked up and tested for FTS and vector
search. Previously the re-rankers were accidentally bypassed in unit
tests, because the builders overrode `.to_arrow()`, but the unit test
called `.to_batches()` which was only defined in the base class. Now all
builders implement `.to_batches()` and leave `.to_arrow()` to the base
class.
4. The `AsyncQueryBase` and `AsyncVectoryQueryBase` setter methods now
return `Self`, which provides the appropriate subclass as the type hint
return value. Previously, `AsyncQueryBase` had them all hard-coded to
`AsyncQuery`, which was unfortunate. (This required bringing in
`typing-extensions` for older Python version, but I think it's worth
it.)
User reported on Discord, when using
`table.vector_search([np.float16(1.0), np.float16(2.0), ...])`, it
yields `TypeError: 'numpy.float16' object is not iterable`
Sometimes it is acceptable to users to only search indexed data and skip
and new un-indexed data. For example, if un-indexed data will be shortly
indexed and they don't mind the delay. In these cases, we can save a lot
of CPU time in search, and provide better latency. Users can activate
this on queries using `fast_search()`.
It's useful to see the underlying query plan for debugging purposes.
This exposes LanceScanner's `explain_plan` function. Addresses #1288
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
SDK
Python
Description
Exposes pyarrow batch api during query execution - relevant when there
is no vector search query, dataset is large and the filtered result is
larger than memory.
---------
Co-authored-by: Ishani Ghose <isghose@amazon.com>
Co-authored-by: Chang She <759245+changhiskhan@users.noreply.github.com>
This changes `lancedb` from a "pure python" setuptools project to a
maturin project and adds a rust lancedb dependency.
The async python client is extremely minimal (only `connect` and
`Connection.table_names` are supported). The purpose of this PR is to
get the infrastructure in place for building out the rest of the async
client.
Although this is not technically a breaking change (no APIs are
changing) it is still a considerable change in the way the wheels are
built because they now include the native shared library.
