Issue with Metadata Filtering on Lambda for LangChain & Pinecone Integration

[vaishnav-mk]

Discussed in #29537

^{Originally posted by vaishnav-mk February 2, 2025}

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

retriever = vector_store.as_retriever(
    search_type="similarity_score_threshold",
    search_kwargs={"k": 5, "score_threshold": 0.5},
)

@tool
def projectTool(query: str) -> str:
    """
    Project tool:
        Handle queries about projects and project details.
    """
    response = retriever.invoke(
        f"{prompts_dict['projectPrompt']}",
        filter={"table_name": "Project", "project_id": projectId},
        k=1,
    )
    logger.info(response)
    return response

Description

I'm using the langchain library in combination with Pinecone for vector retrieval and metadata filtering. The filtering works as expected locally, but when deploying the same code to AWS Lambda, it fails to retrieve the correct data based on the metadata, despite using the same project_id and table_name.

What I’m doing:
I have a vector store retriever that is configured to filter results using metadata, specifically the table_name and project_id. Here’s an example of how I'm invoking the retriever:

retriever = vector_store.as_retriever(
    search_type="similarity_score_threshold",
    search_kwargs={"k": 5, "score_threshold": 0.5},
)

@tool
def projectTool(query: str) -> str:
    """
    Project tool:
        Handle queries about projects and project details.
    Examples:
        - Show me the project details
        - Get the project details
        - Show me the project information
    """
    response = retriever.invoke(
        f"{prompts_dict['projectPrompt']}",
        filter={"table_name": "Project", "project_id": projectId},
        k=1,
    )
    logger.info(response)
    return response

@tool
def materialTool(query: str) -> str:
    """
    Material tool:
        Use ONLY for queries about materials
    Examples:
        - What materials do we have in stock?
        - Show me the material catalog
        - Get material specifications
    DO NOT use for expense-related queries
    """
    response = retriever.invoke(
        f"{prompts_dict['materialPrompt']}", filter={"table_name": "Material", "project_id": projectId}
    )
    logger.info(response)
    return response

What I expect:
The retriever should successfully fetch the project details based on the project_id and table_name metadata filters.

What happens instead:
Locally, everything works as expected, and I can retrieve project data. However, when deploying to Lambda and using the same project_id and table_name, the retrieval fails to return any data.

It works fine with all the other tools except for the project tool for some reason

Data stored in the vector store:
The metadata for the project is stored as follows:

{
  "last_updated": 1738430342,
  "project_id": "18422398-bd99-4228-9337-c0664c9c5045",
  "table_name": "Project",
  "text": "{\"project_summary\": \"This is a detailed summary of the project 'Hello Kitty House'.\\nThe project was initiated on January 29, 2025 and was last updated on January 30, 2025.\\n\\nProject Overview:\\ {...} stages.\"}")}

System Info

System Information
------------------
> OS:  Linux
> OS Version:  #202408141037 SMP PREEMPT_DYNAMIC Wed Aug 14 14:47:07 UTC 2024
> Python Version:  3.12.4 | packaged by Anaconda, Inc. | (main, Jun 18 2024, 15:12:24) [GCC 11.2.0]

Package Information
-------------------
> langchain_core: 0.3.29
> langsmith: 0.2.10
> langchain_cli: 0.0.35
> langserve: 0.3.1

Other Dependencies
------------------
> fastapi: 0.115.6
> gitpython: 3.1.37
> gritql: 0.2.0
> httpx: 0.28.1
> jsonpatch: 1.33
> langserve[all]: Installed. No version info available.
> langsmith-pyo3: Installed. No version info available.
> orjson: 3.10.13
> packaging: 24.0
> pydantic: 2.10.4
> PyYAML: 6.0.1
> requests: 2.32.2
> requests-toolbelt: 1.0.0
> sse-starlette: 1.8.2
> tenacity: 8.2.2
> tomlkit: 0.13.2
> typer[all]: Installed. No version info available.
> typing-extensions: 4.12.2
> uvicorn: 0.34.0
> zstandard: 0.22.0
```</div>

[keenborder786]

It's very difficult to answer your question. From the looks of it, you seem to be passing the filter argument correctly. You can try one thing, rather than passing in the filter duing the invoke, pass the filter when retriever is being initiated:

retriever = vector_store.as_retriever(
    search_type="similarity_score_threshold",
    search_kwargs={"k": 5, "score_threshold": 0.5, "filter": {"table_name": "Material", "project_id": projectId}},
)

[vaishnav-mk]

I tried it earlier, the same thing happens where it works locally but not on AWS lambda

…

On Sun, 2 Feb 2025, 20:32 Mohammad Mohtashim, ***@***.***> wrote: It's very difficult to answer your question. From the looks of it, you seem to be passing the filter argument correctly. You can try one thing, rather than passing in the filter duing the invoke, pass the filter when retriever is being initiated: retriever = vector_store.as_retriever( search_type="similarity_score_threshold", search_kwargs={"k": 5, "score_threshold": 0.5, "filter": {"table_name": "Material", "project_id": projectId}}, ) — Reply to this email directly, view it on GitHub <#29539 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AUE7ZCXNMIYZSQI6DKOI5U32NYXQXAVCNFSM6AAAAABWKFORCKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMMRZGQZTCNJSGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>