- Update ColBertReranker architecture: The current implementation
doesn't use the right arch. This PR uses the implementation in Rerankers
library. Fixes https://github.com/lancedb/lancedb/issues/1546
Benchmark diff (hit rate):
Hybrid - 91 vs 87
reranked vector - 85 vs 80
- Reranking in FTS is basically disabled in main after last week's FTS
updates. I think there's no blocker in supporting that?
- Allow overriding accelerators: Most transformer based Rerankers and
Embedding automatically select device. This PR allows overriding those
settings by passing `device`. Fixes:
https://github.com/lancedb/lancedb/issues/1487
---------
Co-authored-by: BubbleCal <bubble-cal@outlook.com>
Lance now supports FTS, so add it into lancedb Python, TypeScript and
Rust SDKs.
For Python, we still use tantivy based FTS by default because the lance
FTS index now misses some features of tantivy.
For Python:
- Support to create lance based FTS index
- Support to specify columns for full text search (only available for
lance based FTS index)
For TypeScript:
- Change the search method so that it can accept both string and vector
- Support full text search
For Rust
- Support full text search
The others:
- Update the FTS doc
BREAKING CHANGE:
- for Python, this renames the attached score column of FTS from "score"
to "_score", this could be a breaking change for users that rely the
scores
---------
Signed-off-by: BubbleCal <bubble-cal@outlook.com>
Currently targeting the following usage:
```
from lancedb.rerankers import CrossEncoderReranker
reranker = CrossEncoderReranker()
query = "hello"
res1 = table.search(query, vector_column_name="vector").limit(3)
res2 = table.search(query, vector_column_name="text_vector").limit(3)
res3 = table.search(query, vector_column_name="meta_vector").limit(3)
reranked = reranker.rerank_multivector(
[res1, res2, res3],
deduplicate=True,
query=query # some reranker models need query
)
```
- This implements rerank_multivector function in the base reranker so
that all rerankers that implement rerank_vector will automatically have
multivector reranking support
- Special case for RRF reranker that just uses its existing
rerank_hybrid fcn to multi-vector reranking.
---------
Co-authored-by: Weston Pace <weston.pace@gmail.com>
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>
This allows users to specify URIs like:
```
s3+ddb://my_bucket/path?ddbTableName=myCommitTable
```
and it will support concurrent writes in S3.
* [x] Add dynamodb integration tests
* [x] Add modifications to get it working in Python sync API
* [x] Added section in documentation describing how to configure.
Closes#534
---------
Co-authored-by: universalmind303 <cory.grinstead@gmail.com>
Added the ability to specify tokenizer_name, when creating a full text
search index using tantivy. This enables the use of language specific
stemming.
Also updated the [guide on full text
search](https://lancedb.github.io/lancedb/fts/) with a short section on
choosing tokenizer.
Fixes#1315
- Tried to address some onboarding feedbacks listed in
https://github.com/lancedb/lancedb/issues/1224
- Improve visibility of pydantic integration and embedding API. (Based
on onboarding feedback - Many ways of ingesting data, defining schema
but not sure what to use in a specific use-case)
- Add a guide that takes users through testing and improving retriever
performance using built-in utilities like hybrid-search and reranking
- Add some benchmarks for the above
- Add missing cohere docs
---------
Co-authored-by: Weston Pace <weston.pace@gmail.com>
- add `return` for `__enter__`
The buggy code didn't return the object, therefore it will always return
None within a context manager:
```python
with await lancedb.connect_async("./.lancedb") as db:
# db is always None
```
(BTW, why not to design an async context manager?)
- add a unit test for Async connection context manager
- update return type of `AsyncConnection.open_table` to `AsyncTable`
Although type annotation doesn't affect the functionality, it is helpful
for IDEs.
The optimize function is pretty crucial for getting good performance
when building a large scale dataset but it was only exposed in rust
(many sync python users are probably doing this via to_lance today)
This PR adds the optimize function to nodejs and to python.
I left the function marked experimental because I think there will
likely be changes to optimization (e.g. if we add features like
"optimize on write"). I also only exposed the `cleanup_older_than`
configuration parameter since this one is very commonly used and the
rest have sensible defaults and we don't really know why we would
recommend different values for these defaults anyways.
This was already configurable in the rust API but it wasn't actually
being passed down to the underlying dataset. I added this option to both
the async python API and the new nodejs API.
I also added this option to the synchronous python API.
I did not add the option to vectordb.
closes#1194#1172#1124#1208
@wjones127 : `if query_type != "fts":` is needed because both fts and
vector search create `LanceQueryBuilder` which has `vector_column_name`
as a required attribute.
Exposes `storage_options` in LanceDB. This is provided for Python async,
Node `lancedb`, and Node `vectordb` (and Rust of course). Python
synchronous is omitted because it's not compatible with the PyArrow
filesystems we use there currently. In the future, we will move the sync
API to wrap the async one, and then it will get support for
`storage_options`.
1. Fixes#1168
2. Closes#1165
3. Closes#1082
4. Closes#439
5. Closes#897
6. Closes#642
7. Closes#281
8. Closes#114
9. Closes#990
10. Deprecating `awsCredentials` and `awsRegion`. Users are encouraged
to use `storageOptions` instead.
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 PR adds support for passing through a set of ordering fields at
index time (unsigned ints that tantivity can use as fast_fields) that at
query time you can sort your results on. This is useful for cases where
you want to get related hits, i.e by keyword, but order those hits by
some other score, such as popularity.
I.e search for songs descriptions that match on "sad AND jazz AND 1920"
and then order those by number of times played. Example usage can be
seen in the fts tests.
---------
Co-authored-by: Nat Roth <natroth@Nats-MacBook-Pro.local>
Co-authored-by: Chang She <759245+changhiskhan@users.noreply.github.com>
HuggingFace Dataset is written as arrow batches.
For DatasetDict, all splits are written with a "split" column appended.
- [x] what if the dataset schema already has a `split` column
- [x] add unit tests
solves https://github.com/lancedb/lancedb/issues/1086
Usage Reranking with FTS:
```
retriever = db.create_table("fine-tuning", schema=Schema, mode="overwrite")
pylist = [{"text": "Carson City is the capital city of the American state of Nevada. At the 2010 United States Census, Carson City had a population of 55,274."},
{"text": "The Commonwealth of the Northern Mariana Islands is a group of islands in the Pacific Ocean that are a political division controlled by the United States. Its capital is Saipan."},
{"text": "Charlotte Amalie is the capital and largest city of the United States Virgin Islands. It has about 20,000 people. The city is on the island of Saint Thomas."},
{"text": "Washington, D.C. (also known as simply Washington or D.C., and officially as the District of Columbia) is the capital of the United States. It is a federal district. "},
{"text": "Capital punishment (the death penalty) has existed in the United States since before the United States was a country. As of 2017, capital punishment is legal in 30 of the 50 states."},
{"text": "North Dakota is a state in the United States. 672,591 people lived in North Dakota in the year 2010. The capital and seat of government is Bismarck."},
]
retriever.add(pylist)
retriever.create_fts_index("text", replace=True)
query = "What is the capital of the United States?"
reranker = CohereReranker(return_score="all")
print(retriever.search(query, query_type="fts").limit(10).to_pandas())
print(retriever.search(query, query_type="fts").rerank(reranker=reranker).limit(10).to_pandas())
```
Result
```
text vector score
0 Capital punishment (the death penalty) has exi... [0.099975586, 0.047943115, -0.16723633, -0.183... 0.729602
1 Charlotte Amalie is the capital and largest ci... [-0.021255493, 0.03363037, -0.027450562, -0.17... 0.678046
2 The Commonwealth of the Northern Mariana Islan... [0.3684082, 0.30493164, 0.004600525, -0.049407... 0.671521
3 Carson City is the capital city of the America... [0.13989258, 0.14990234, 0.14172363, 0.0546569... 0.667898
4 Washington, D.C. (also known as simply Washing... [-0.0090408325, 0.42578125, 0.3798828, -0.3574... 0.653422
5 North Dakota is a state in the United States. ... [0.55859375, -0.2109375, 0.14526367, 0.1634521... 0.639346
text vector score _relevance_score
0 Washington, D.C. (also known as simply Washing... [-0.0090408325, 0.42578125, 0.3798828, -0.3574... 0.653422 0.979977
1 The Commonwealth of the Northern Mariana Islan... [0.3684082, 0.30493164, 0.004600525, -0.049407... 0.671521 0.299105
2 Capital punishment (the death penalty) has exi... [0.099975586, 0.047943115, -0.16723633, -0.183... 0.729602 0.284874
3 Carson City is the capital city of the America... [0.13989258, 0.14990234, 0.14172363, 0.0546569... 0.667898 0.089614
4 North Dakota is a state in the United States. ... [0.55859375, -0.2109375, 0.14526367, 0.1634521... 0.639346 0.063832
5 Charlotte Amalie is the capital and largest ci... [-0.021255493, 0.03363037, -0.027450562, -0.17... 0.678046 0.041462
```
## Vector Search usage:
```
query = "What is the capital of the United States?"
reranker = CohereReranker(return_score="all")
print(retriever.search(query).limit(10).to_pandas())
print(retriever.search(query).rerank(reranker=reranker, query=query).limit(10).to_pandas()) # <-- Note: passing extra string query here
```
Results
```
text vector _distance
0 Capital punishment (the death penalty) has exi... [0.099975586, 0.047943115, -0.16723633, -0.183... 39.728973
1 Washington, D.C. (also known as simply Washing... [-0.0090408325, 0.42578125, 0.3798828, -0.3574... 41.384884
2 Carson City is the capital city of the America... [0.13989258, 0.14990234, 0.14172363, 0.0546569... 55.220200
3 Charlotte Amalie is the capital and largest ci... [-0.021255493, 0.03363037, -0.027450562, -0.17... 58.345654
4 The Commonwealth of the Northern Mariana Islan... [0.3684082, 0.30493164, 0.004600525, -0.049407... 60.060867
5 North Dakota is a state in the United States. ... [0.55859375, -0.2109375, 0.14526367, 0.1634521... 64.260544
text vector _distance _relevance_score
0 Washington, D.C. (also known as simply Washing... [-0.0090408325, 0.42578125, 0.3798828, -0.3574... 41.384884 0.979977
1 The Commonwealth of the Northern Mariana Islan... [0.3684082, 0.30493164, 0.004600525, -0.049407... 60.060867 0.299105
2 Capital punishment (the death penalty) has exi... [0.099975586, 0.047943115, -0.16723633, -0.183... 39.728973 0.284874
3 Carson City is the capital city of the America... [0.13989258, 0.14990234, 0.14172363, 0.0546569... 55.220200 0.089614
4 North Dakota is a state in the United States. ... [0.55859375, -0.2109375, 0.14526367, 0.1634521... 64.260544 0.063832
5 Charlotte Amalie is the capital and largest ci... [-0.021255493, 0.03363037, -0.027450562, -0.17... 58.345654 0.041462
```
I know there's a larger effort to have the python client based on the
core rust implementation, but in the meantime there have been several
issues (#1072 and #485) with some of the azure blob storage calls due to
pyarrow not natively supporting an azure backend. To this end, I've
added an optional import of the fsspec implementation of azure blob
storage [`adlfs`](https://pypi.org/project/adlfs/) and passed it to
`pyarrow.fs`. I've modified the existing test and manually verified it
with some real credentials to make sure it behaves as expected.
It should be now as simple as:
```python
import lancedb
db = lancedb.connect("az://blob_name/path")
table = db.open_table("test")
table.search(...)
```
Thank you for this cool project and we're excited to start using this
for real shortly! 🎉 And thanks to @dwhitena for bringing it to my
attention with his prediction guard posts.
Co-authored-by: christiandilorenzo <christian.dilorenzo@infiniaml.com>
The LanceDB embeddings registry allows users to annotate the pydantic
model used as table schema with the desired embedding function, e.g.:
```python
class Schema(LanceModel):
id: str
vector: Vector(openai.ndims()) = openai.VectorField()
text: str = openai.SourceField()
```
Tables created like this does not require embeddings to be calculated by
the user explicitly, e.g. this works:
```python
table.add([{"id": "foo", "text": "rust all the things"}])
```
However, trying to construct pydantic model instances without vector
doesn't because it's a required field.
Instead, you need add a default value:
```python
class Schema(LanceModel):
id: str
vector: Vector(openai.ndims()) = openai.VectorField(default=None)
text: str = openai.SourceField()
```
then this completes without errors:
```python
table.add([Schema(id="foo", text="rust all the things")])
```
However, all of the vectors are filled with zeros. Instead in
add_vector_col we have to add an additional check so that the embedding
generation is called.
The synchronous table_names function in python lancedb relies on arrow's
filesystem which behaves slightly differently than object_store. As a
result, the function would not work properly in GCS.
However, the async table_names function uses object_store directly and
thus is accurate. In most cases we can fallback to using the async
table_names function and so this PR does so. The one case we cannot is
if the user is already in an async context (we can't start a new async
event loop). Soon, we can just redirect those users to use the async API
instead of the sync API and so that case will eventually go away. For
now, we fallback to the old behavior.
1. filtering with fts mutated the schema, which caused schema mistmatch
problems with hybrid search as it combines fts and vector search tables.
2. fts with filter failed with `with_row_id`. This was because row_id
was calculated before filtering which caused size mismatch on attaching
it after.
3. The fix for 1 meant that now row_id is attached before filtering but
passing a filter to `to_lance` on a dataset that already contains
`_rowid` raises a panic from lance. So temporarily, in case where fts is
used with a filter AND `with_row_id`, we just force user to using the
duckdb pathway.
---------
Co-authored-by: Chang She <759245+changhiskhan@users.noreply.github.com>
In order to add support for `add` we needed to migrate the rust `Table`
trait to a `Table` struct and `TableInternal` trait (similar to the way
the connection is designed).
While doing this we also cleaned up some inconsistencies between the
SDKs:
* Python and Node are garbage collected languages and it can be
difficult to trigger something to be freed. The convention for these
languages is to have some kind of close method. I added a close method
to both the table and connection which will drop the underlying rust
object.
* We made significant improvements to table creation in
cc5f2136a6
for the `node` SDK. I copied these changes to the `nodejs` SDK.
* The nodejs tables were using fs to create tmp directories and these
were not getting cleaned up. This is mostly harmless but annoying and so
I changed it up a bit to ensure we cleanup tmp directories.
* ~~countRows in the node SDK was returning `bigint`. I changed it to
return `number`~~ (this actually happened in a previous PR)
* Tables and connections now implement `std::fmt::Display` which is
hooked into python's `__repr__`. Node has no concept of a regular "to
string" function and so I added a `display` method.
* Python method signatures are changing so that optional parameters are
always `Optional[foo] = None` instead of something like `foo = False`.
This is because we want those defaults to be in rust whenever possible
(though we still need to mention the default in documentation).
* I changed the python `AsyncConnection/AsyncTable` classes from
abstract classes with a single implementation to just classes because we
no longer have the remote implementation in python.
Note: this does NOT add the `add` function to the remote table. This PR
was already large enough, and the remote implementation is unique
enough, that I am going to do all the remote stuff at a later date (we
should have the structure in place and correct so there shouldn't be any
refactor concerns)
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
