Skip to content

DuckSearch

Efficient BM25 with DuckDB 🦆

documentation license

DuckSearch is a lightweight and easy-to-use library to search documents. DuckSearch is built on top of DuckDB, a high-performance analytical database. DuckDB is designed to execute analytical SQL queries fast, and DuckSearch leverages this to provide efficient search and filtering features. DuckSearch index can be updated with new documents and documents can be deleted as well. DuckSearch also supports HuggingFace datasets, allowing to index datasets directly from the HuggingFace Hub.

Installation

Install DuckSearch using pip:

pip install ducksearch

Documentation

The complete documentation is available here, which includes in-depth guides, examples, and API references.

Upload

We can upload documents to DuckDB using the upload.documents function. The documents are stored in a DuckDB database, and the fields are indexed with BM25.

from ducksearch import upload

documents = [
    {
        "id": 0,
        "title": "Hotel California",
        "style": "rock",
        "date": "1977-02-22",
        "popularity": 9,
    },
    {
        "id": 1,
        "title": "Here Comes the Sun",
        "style": "rock",
        "date": "1969-06-10",
        "popularity": 10,
    },
    {
        "id": 2,
        "title": "Alive",
        "style": "electro, punk",
        "date": "2007-11-19",
        "popularity": 9,
    },
]

upload.documents(
    database="ducksearch.duckdb",
    key="id", # Unique document identifier
    fields=["title", "style"], # List of fields to use for search.
    documents=documents,
    dtypes={
        "date": "DATE",
        "popularity": "INT",
    },
)

search.documents returns a list of list of documents ordered by relevance. We can control the number of documents to return using the top_k parameter. The following example demonstrates how to search for documents with the queries "punk" and "california" while filtering the results to include only documents with a date after 1970 and a popularity score greater than 8. We will order the results by a weighted sum of the BM25 score and the popularity score provided in the document.

from ducksearch import search

search.documents(
    database="ducksearch.duckdb",
    queries=["punk", "california"],
    top_k=10,
    filters="YEAR(date) >= 1970 AND popularity > 8",
    order_by="0.8 * score + 0.2 * popularity DESC",
)
[
    [
        {
            "id": "2",
            "title": "Alive",
            "style": "electro, punk",
            "date": Timestamp("2007-11-19 00:00:00"),
            "popularity": 9,
            "score": 0.17841622233390808,
        }
    ],
    [
        {
            "id": "0",
            "title": "Hotel California",
            "style": "rock, pop",
            "date": Timestamp("1977-02-22 00:00:00"),
            "popularity": 9,
            "score": 0.156318798661232,
        }
    ],
]

Filters are SQL expressions that are applied to the search results. We can use every filtering function DuckDB provides such as date functions.

Both filters and order_by parameters are optional. If not provided, the results are ordered by BM25 relevance and no filters are applied.

Delete and update index

We can delete documents and update the BM25 weights accordingly using the delete.documents function.

from ducksearch import delete

delete.documents(
    database="ducksearch.duckdb",
    ids=[0, 1],
)

To update the index, we should first delete the documents and then upload the updated documents.

Extra features

HuggingFace

The upload.documents function can also index HuggingFace datasets directly from the url. The following example demonstrates how to index the FineWeb dataset from HuggingFace. We will use the fields "text" and "url" for search. We will also specify the data types for the "date", "token_count", and "language_score" fields to be able to filter the results.

from ducksearch import upload

upload.documents(
    database="fineweb.duckdb",
    key="id",
    fields=["text", "url"],
    documents="https://huggingface.co/datasets/HuggingFaceFW/fineweb/resolve/main/sample/10BT/000_00000.parquet",
    dtypes={
        "date": "DATE",
        "token_count": "INT",
        "language_score": "FLOAT",
    },
    limit=3000, # demonstrate with a small dataset
)

We can then search the FineWeb dataset with the search.documents function. We order the results by BM25 score and then date.

from ducksearch import search

search.documents(
    database="fineweb.duckdb",
    queries=["earth science"],
    top_k=2,
    order_by="score DESC, date DESC",
)
[
    [
        {
            "id": "<urn:uuid:1e6ae53b-e0d7-431b-8d46-290244e597e9>",
            "text": "Earth Science Tutors in Rowland...",
            "id_1": "<urn:uuid:1e6ae53b-e0d7-431b-8d46-290244e597e9>",
            "dump": "CC-MAIN-2017-34",
            "url": "http://rowland.universitytutor.com/rowland_earth-science-tutoring",
            "date": Timestamp("2017-08-19 00:00:00"),
            "file_path": "s3://commoncrawl/crawl-data/CC-MAIN-2017-34/segments/1502886105304.35/warc/CC-MAIN-20170819051034-20170819071034-00240.warc.gz",
            "language": "en",
            "language_score": 0.8718525171279907,
            "token_count": 313,
            "bm25id": 523,
            "score": 2.3761106729507446,
        },
        {
            "id": "<urn:uuid:cd94a04f-1632-4c8b-81d2-cb353163116e>",
            "text": "- Geomagnetic field....",
            "id_1": "<urn:uuid:cd94a04f-1632-4c8b-81d2-cb353163116e>",
            "dump": "CC-MAIN-2022-21",
            "url": "https://www.imperial.ac.uk/people/adrian.muxworthy/?respub-action=citation.html&id=1149861&noscript=noscript",
            "date": Timestamp("2022-05-20 00:00:00"),
            "file_path": "s3://commoncrawl/crawl-data/CC-MAIN-2022-21/segments/1652662530553.34/warc/CC-MAIN-20220519235259-20220520025259-00601.warc.gz",
            "language": "en",
            "language_score": 0.8225595951080322,
            "token_count": 517,
            "bm25id": 4783,
            "score": 2.3569871187210083,
        },
    ]
]

Note: by default, results are ordered by BM25 relevance.

Tables

Ducksearch creates two distinct schemas: bm25_tables, bm25_documents.

  • We can find the uploaded documents in the bm25_tables.documents table.

  • We can find the inverted index in the bm25_documents.scores table. You can update the scores as you wish. Just note that tokens scores will be updated each time you upload documents (every tokens scores mentionned in the set of uploaded documents).

  • We can update the set of stopwords in the bm25_documents.stopwords table.

Benchmark

Dataset ndcg@10 hits@1 hits@10 mrr@10 map@10 r-precision qps Indexation Time (s) Number of Documents and Queries
arguana 0.3779 0.0 0.8267 0.2491 0.2528 0.0108 117.80 1.42 1,406 queries, 8.67K documents
climate-fever 0.1184 0.1068 0.3648 0.1644 0.0803 0.0758 5.88 302.39 1,535 queries, 5.42M documents
dbpedia-entity 0.6046 0.7669 5.6241 0.8311 0.0649 0.0741 113.20 181.42 400 queries, 4.63M documents
fever 0.3861 0.2583 0.5826 0.3525 0.3329 0.2497 74.40 329.70 6,666 queries, 5.42M documents
fiqa 0.2445 0.2207 0.6790 0.3002 0.1848 0.1594 545.77 6.04 648 queries, 57K documents
hotpotqa 0.4487 0.5059 0.9699 0.5846 0.3642 0.3388 48.15 163.14 7,405 queries, 5.23M documents
msmarco 0.8951 1.0 8.6279 1.0 0.0459 0.0473 35.11 202.37 6,980 queries, 8.84M documents
nfcorpus 0.3301 0.4396 2.4087 0.5292 0.1233 0.1383 3464.66 0.99 323 queries, 3.6K documents
nq 0.2451 0.1272 0.4574 0.2099 0.1934 0.1240 150.23 71.43 3,452 queries, 2.68M documents
quora 0.7705 0.6783 1.1749 0.7606 0.7206 0.6502 741.13 3.78 10,000 queries, 523K documents
scidocs 0.1025 0.1790 0.8240 0.2754 0.0154 0.0275 879.11 4.46 1,000 queries, 25K documents
scifact 0.6908 0.5533 0.9133 0.6527 0.6416 0.5468 2153.64 1.22 300 queries, 5K documents
trec-covid 0.9533 1.0 9.4800 1.0 0.0074 0.0077 112.38 22.15 50 queries, 171K documents
webis-touche2020 0.4130 0.5510 3.7347 0.7114 0.0564 0.0827 104.65 44.14 49 queries, 382K documents

References

  • DuckDB

  • DuckDB Full Text Search: Note that DuckSearch rely partially on the DuckDB Full Text Search extension but accelerate the search process via top_k_token approximation, pre-computation of scores and multi-threading.

License

DuckSearch is released under the MIT license.

Citation

@misc{PyLate,
  title={DuckSearch, efficient search with DuckDB},
  author={Sourty, Raphael},
  url={https://github.com/lightonai/ducksearch},
  year={2024}
}