The basic idea of MRR is this -
https://www.evidentlyai.com/ranking-metrics/mean-reciprocal-rank-mrr
I've implemented a weighted version for allowing user to set weightage
between vector and fts.
The gist is something like this
### Scenario A: Document at rank 1 in one set, absent from another
```
# Assuming equal weights: weight_vector = 0.5, weight_fts = 0.5
vector_rr = 1.0 # rank 1 → 1/1 = 1.0
fts_rr = 0.0 # absent → 0.0
weighted_mrr = 0.5 × 1.0 + 0.5 × 0.0 = 0.5
```
### Scenario B: Document at rank 1 in one set, rank 2 in another
```
# Same weights: weight_vector = 0.5, weight_fts = 0.5
vector_rr = 1.0 # rank 1 → 1/1 = 1.0
fts_rr = 0.5 # rank 2 → 1/2 = 0.5
weighted_mrr = 0.5 × 1.0 + 0.5 × 0.5 = 0.5 + 0.25 = 0.75
```
And so with `return_score="all"` the result looks something like this
(this is from the reranker tests).
Because this is a weighted rank based reranker, some results might have
the same score
```
text vector _distance _rowid _score _relevance_score
0 I am your father [-0.010703234, 0.069315575, 0.030076642, 0.002... 8.149148e-13 8589934598 10.978719 1.000000
1 the ground beneath my feet [-0.09500901, 0.00092102867, 0.0755851, 0.0372... 1.376896e+00 8589934604 NaN 0.250000
2 I find your lack of faith disturbing [0.07525753, -0.0100010475, 0.09990541, 0.0209... NaN 8589934595 3.483394 0.250000
3 but I don't wanna die [0.033476487, -0.011235877, -0.057625435, -0.0... 1.538222e+00 8589934610 1.130355 0.238095
4 if you strike me down I shall become more powe... [0.00432201, 0.030120496, 5.3317923e-05, 0.033... 1.381086e+00 8589934594 0.715157 0.216667
5 I see a salty message written in the eves [-0.04213107, 0.0016004723, 0.061052393, -0.02... 1.638301e+00 8589934603 1.043785 0.133333
6 but his son was mortal [0.012462767, 0.049041674, -0.057339743, -0.04... 1.421566e+00 8589934620 NaN 0.125000
7 I've got a bad feeling about this [-0.06973199, -0.029960092, 0.02641632, -0.031... NaN 8589934596 1.043785 0.125000
8 now that's a name I haven't heard in a long time [-0.014374257, -0.013588792, -0.07487557, 0.03... 1.597573e+00 8589934593 0.848772 0.118056
9 he was a god [-0.0258895, 0.11925236, -0.029397793, 0.05888... 1.423147e+00 8589934618 NaN 0.100000
10 I wish they would make another one [-0.14737535, -0.015304729, 0.04318139, -0.061... NaN 8589934622 1.043785 0.100000
11 Kratos had a son [-0.057455737, 0.13734367, -0.03537109, -0.000... 1.488075e+00 8589934617 NaN 0.083333
12 I don't wanna live like this [-0.0028891307, 0.015214227, 0.025183653, 0.08... NaN 8589934609 1.043785 0.071429
13 I see a mansard roof through the trees [0.052383978, 0.087759204, 0.014739997, 0.0239... NaN 8589934602 1.043785 0.062500
14 great kid don't get cocky [-0.047043696, 0.054648954, -0.008509666, -0.0... 1.618125e+00 8589934592 NaN 0.055556
```
Previously `return_score="all"` was supported only for the default
reranker (RRF) and not the model based rerankers.
This adds support for keeping all scores in the base reranker so that
all model based rerankers can use it. Its a slower path than keeping
just the relevance score but can be useful in debugging
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **Documentation**
- Added an integration banner image to the beginning of the
Genkitx-LanceDB documentation.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **Documentation**
- Added a comprehensive guide for integrating LanceDB with Genkit,
including installation instructions, setup, indexing, retrieval, and
building a custom RAG pipeline with example code and screenshots.
- Updated the documentation navigation to include the new Genkit
integration, making it accessible from the site menu.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
I found some edge cases while running experiments that - depending on
the base reranking libraries, some of them don't handle empty lists
well. This PR manually checks if the result set to be reranked is empty
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **Bug Fixes**
- Enhanced search result processing by ensuring that reordering only
occurs when valid, non-empty results are available, thereby preventing
unnecessary operations and potential errors.
- **Tests**
- Added automated tests to verify that empty search result sets are
handled correctly, ensuring consistent behavior across various
rerankers.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
- fixes https://github.com/lancedb/lancedb/issues/1697.
- unifies vector column inference logic for remote and local table to
prevent future disparities.
- Updates docstring in RemoteTable to specify empty queries are not
supported
Right now when passing vector and query explicitly for hybrid search ,
vector_column_name is not deduced.
(https://lancedb.github.io/lancedb/hybrid_search/hybrid_search/#hybrid-search-in-lancedb
). Because vector and query can be both none when initialising the
QueryBuilder in this case. This PR forces deduction of query type if it
is set to "hybrid"
- Enforce all rerankers always return _relevance_score. This was already
loosely done in tests before but based on user feedback its better to
always have _relevance_score present in all reranked results
- Deprecate LinearCombinationReranker in docs. And also fix a case where
it would not return _relevance_score if one result set was missing
- Both LinearCombination (the current default) and RRF are pretty fast
compared to model based rerankers. RRF is slightly faster.
- In our tests RRF has also been slightly more accurate.
This PR:
- Makes RRF the default reranker
- Removed duplicate docs for rerankers
Currently, the only documented way of performing hybrid search is by
using embedding API and passing string queries that get automatically
embedded. There are use cases where users might like to pass vectors and
text manually instead.
This ticket contains more information and historical context -
https://github.com/lancedb/lancedb/issues/937
This breaks a undocumented pathway that allowed passing (vector, text)
tuple queries which was intended to be temporary, so this is marked as a
breaking change. For all practical purposes, this should not really
impact most users
### usage
```
results = table.search(query_type="hybrid")
.vector(vector_query)
.text(text_query)
.limit(5)
.to_pandas()
```
This PR:
- Adds missing license headers
- Integrates with answerdotai Rerankers package
- Updates ColbertReranker to subclass answerdotai package. This is done
to keep backwards compatibility as some users might be used to importing
ColbertReranker directly
- Set `trust_remote_code` to ` True` by default in CrossEncoder and
sentence-transformer based rerankers
- 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>
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>
- 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>
https://github.com/lancedb/lancedb/issues/1266#event-12703166915
This happens because openai API errors out with None values. The current
log level didn't really print out the msg on screen. Changed the log
level to warning, which better suits this case.
Also, retry loop can be disabled by setting `max_retries=0` (I'm not
sure if we should also set this as the default behaviour as hitting api
rate is quite common when ingesting large corpus)
```
func = get_registry().get("openai").create(max_retries=0)
````
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
```
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>
I think this should work. Need to deploy it to be sure as it can be
tested locally. Can be tested here.
2 things about this solution:
* All pages have a same meta tag, i.e, lancedb banner
* If needed, we can automatically use the first image of each page and
generate meta tags using the ultralytics mkdocs plugin that we did for
this purpose - https://github.com/ultralytics/mkdocs
- Rename safe_import -> attempt_import_or_raise (closes
https://github.com/lancedb/lancedb/pull/923)
- Update docs
- Add Notebook example (@changhiskhan you can use it for the talk. Comes
with "open in colab" button)
- Latency benchmark & results comparison, sanity check on real-world
data
- Updates the default openai model to gpt-4
based on https://github.com/lancedb/lancedb/pull/713
- The Reranker api can be plugged into vector only or fts only search
but this PR doesn't do that (see example -
https://txt.cohere.com/rerank/)
### Default reranker -- `LinearCombinationReranker(weight=0.7,
fill=1.0)`
```
table.search("hello", query_type="hybrid").rerank(normalize="score").to_pandas()
```
### Available rerankers
LinearCombinationReranker
```
from lancedb.rerankers import LinearCombinationReranker
# Same as default
table.search("hello", query_type="hybrid").rerank(
normalize="score",
reranker=LinearCombinationReranker()
).to_pandas()
# with custom params
reranker = LinearCombinationReranker(weight=0.3, fill=1.0)
table.search("hello", query_type="hybrid").rerank(
normalize="score",
reranker=reranker
).to_pandas()
```
Cohere Reranker
```
from lancedb.rerankers import CohereReranker
# default model.. English and multi-lingual supported. See docstring for available custom params
table.search("hello", query_type="hybrid").rerank(
normalize="rank", # score or rank
reranker=CohereReranker()
).to_pandas()
```
CrossEncoderReranker
```
from lancedb.rerankers import CrossEncoderReranker
table.search("hello", query_type="hybrid").rerank(
normalize="rank",
reranker=CrossEncoderReranker()
).to_pandas()
```
## Using custom Reranker
```
from lancedb.reranker import Reranker
class CustomReranker(Reranker):
def rerank_hybrid(self, vector_result, fts_result):
combined_res = self.merge_results(vector_results, fts_results) # or use custom combination logic
# Custom rerank logic here
return combined_res
```
- [x] Expand testing
- [x] Make sure usage makes sense
- [x] Run simple benchmarks for correctness (Seeing weird result from
cohere reranker in the toy example)
- Support diverse rerankers by default:
- [x] Cross encoding
- [x] Cohere
- [x] Reciprocal Rank Fusion
---------
Co-authored-by: Chang She <759245+changhiskhan@users.noreply.github.com>
Co-authored-by: Prashanth Rao <35005448+prrao87@users.noreply.github.com>
Named it Gemini-text for now. Not sure how complicated it will be to
support both text and multimodal embeddings under the same class
"gemini"..But its not something to worry about for now I guess.
- Register open_table as event
- Because we're dropping 'seach' event currently, changed the name to
'search_table' and introduced throttling
- Throttled events will be counted once per time batch so that the user
is registered but event count doesn't go up by a lot
* Filename typo
* Remove rick_morty csv as users won't really be able to use it.. We can
create a an executable colab and download it from a bucket or smth.
Sets things up for this -> https://github.com/lancedb/lancedb/issues/579
- Just separates out the registry/ingestion code from the function
implementation code
- adds a `get_registry` util
- package name "open-clip" -> "open-clip-torch"
This PR adds an overview of embeddings docs:
- 2 ways to vectorize your data using lancedb - explicit & implicit
- explicit - manually vectorize your data using `wit_embedding` function
- Implicit - automatically vectorize your data as it comes by ingesting
your embedding function details as table metadata
- Multi-modal example w/ disappearing embedding function
It improves the UX as iterators can be of any type supported by the
table (plus recordbatch) & there is no separate requirement.
Also expands the test cases for pydantic & arrow schema.
If this is looks good I'll update the docs.
Example usage:
```
class Content(LanceModel):
vector: vector(2)
item: str
price: float
def make_batches():
for _ in range(5):
yield from [
# pandas
pd.DataFrame({
"vector": [[3.1, 4.1], [1, 1]],
"item": ["foo", "bar"],
"price": [10.0, 20.0],
}),
# pylist
[
{"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
{"vector": [5.9, 26.5], "item": "bar", "price": 20.0},
],
# recordbatch
pa.RecordBatch.from_arrays(
[
pa.array([[3.1, 4.1], [5.9, 26.5]], pa.list_(pa.float32(), 2)),
pa.array(["foo", "bar"]),
pa.array([10.0, 20.0]),
],
["vector", "item", "price"],
),
# pydantic list
[
Content(vector=[3.1, 4.1], item="foo", price=10.0),
Content(vector=[5.9, 26.5], item="bar", price=20.0),
]]
db = lancedb.connect("db")
tbl = db.create_table("tabley", make_batches(), schema=Content, mode="overwrite")
tbl.add(make_batches())
```
Same should with arrow schema.
---------
Co-authored-by: Weston Pace <weston.pace@gmail.com>
* Rename "Reference" -> "Guides" to create distinction b/w api reference
and user facing docs
* Add all the various ways to create, add and delete from table
Related - https://github.com/lancedb/lancedb/pull/391
Makes the following work so all the formats accepted by `create_table()`
are also accepted by `add()`
```
import lancedb
import pyarrow as pa
db = lancedb.connect("/tmp")
def make_batches():
for i in range(5):
yield pa.RecordBatch.from_arrays(
[
pa.array([[3.1, 4.1], [5.9, 26.5]]),
pa.array(["foo", "bar"]),
pa.array([10.0, 20.0]),
],
["vector", "item", "price"],
)
schema = pa.schema([
pa.field("vector", pa.list_(pa.float32())),
pa.field("item", pa.utf8()),
pa.field("price", pa.float32()),
])
tbl = db.create_table("table4", make_batches(), schema=schema)
tbl.add(make_batches())
```
Ayush Chaurasia
