community: added the possibility to return document ids as part of the results of similarity search in vectorstore chroma

[majdabd]

Hello, I modified the _results_to_docs_and_scores method in chroma.py to allow the results of the similarity search to include the IDs of documents or document chunks. I'm new to Chroma and LangChain, but I couldn't find a way to retrieve the IDs of documents that best match the query.

The new _results_to_docs_and_scores thus becomes:

def _results_to_docs_and_scores(results: Any) -> List[Tuple[Document, float]]:
    return [
        # TODO: Chroma can do batch querying,
        # we shouldn't hard code to the 1st result
        (Document(page_content=result[0], metadata=(result[1] | {'id': result[3]}) or {}), result[2])
        for result in zip(
            results["documents"][0],
            results["metadatas"][0],
            results["distances"][0],
            results["ids"][0]
        )
    ]

Note: I found this solution in this closed issue #11592, but as far as I know it has not been proposed in a pull request.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		Mar 29, 2024 0:14am

[baskaryan]

cc @jeffchuber

[jeffchuber]

this looks like if you have metadata it will override an id prop... can we think of a different way to handle this? i think it may really confuse some users, albeit an edge case

[TheDeafOne]

@hwchase17 why was this closed? Was there a solution merged? Having this functionality would be super useful.

[LDelPinoNT]

From the messages I understand there is a variable called id, but it can be useful to have the document id, maybe as "doc_id" (so there is no name overlapping).

An example of use case for the "doc_id" is recover it for training and evaluating RAG with InformationRetrievalEvaluator of the package SentenceEmbeddings (https://sbert.net/docs/package_reference/sentence_transformer/evaluation.html#informationretrievalevaluator) This evaluator needs pairs of documents and documents ids as "corpus" parameter which can be retrieved directly from VectorStores.