Filter doesn't work, when I use prisma vector store as a retriever

[ykshev]

I'd like to use prisma vectror store as a memory retrieval and have an ability to filter the messages by UserId, like this await vectorStore.similaritySearchWithScore(input, 3, { userId: { equals: userId } }). When I'm trying to pass such a filter to asRetriever method, I see no success, what I'm doing wrong?

const vectorStore = PrismaVectorStore.withModel<Message>(prisma).create(
  new OpenAIEmbeddings(),
  {
    prisma: Prisma,
    tableName: "Message",
    vectorColumnName: "vector",
    columns: {
      id: PrismaVectorStore.IdColumn,
      content: PrismaVectorStore.ContentColumn,
    },
  }
);
const memory = new VectorStoreRetrieverMemory({
  vectorStoreRetriever: vectorStore.asRetriever(1), // How I can pass a filter query here?
  memoryKey: "history",
});

const chain = new ConversationChain({
  memory,
  prompt,
  llm,
})

[dosubot[bot]]

🤖

Hello,

Based on the information you've provided, it seems you're trying to pass a filter to the asRetriever method. According to a similar issue in the LangChainJS repository, the asRetriever method accepts two optional parameters: k and filter. The filter parameter is where you can pass your metadata filters. Here's how you can do it:

const filter = { 'userId': { equals: userId } };
const retriever = vectorStore.asRetriever(undefined, filter);

In this example, filter is an object that contains the metadata filters you want to apply. The asRetriever method is then called on the vectorStore object, passing filter as the second argument. The first argument is undefined because we are not specifying a value for k in this case. The method returns a VectorStoreRetriever object that you can use to retrieve documents based on the specified filters.

However, in your case, you're using the VectorStoreRetrieverMemory class, which internally uses the asRetriever method. To pass a filter query to the VectorStoreRetrieverMemory class, you can do so by providing the filter as part of the searchParams object in the options passed to the fromLLM method.

Here is an example of how you can do this:

const selfQueryRetriever = await SelfQueryRetriever.fromLLM({
  llm,
  vectorStore,
  documentContents,
  attributeInfo,
  structuredQueryTranslator: new FunctionalTranslator(),
  searchParams: {
    filter: (doc: Document) => doc.metadata && doc.metadata.userId === userId,
    mergeFiltersOperator: "or",
    k: docs.length,
  },
});

In this example, the filter function includes only documents where the userId metadata field matches the userId you're looking for. The mergeFiltersOperator key in the searchParams object determines how this filter is combined with any other filters that are applied. If mergeFiltersOperator is "or", then a document is included if it passes either the filter function or the other filters. If mergeFiltersOperator is "and", then a document is included only if it passes both the filter function and the other filters.

I hope this helps! Let me know if you have any other questions.

Sources

• missing search kwargs for asRetriever() method for vectorstores.
• langchain/src/vectorstores/base.ts
• langchain/src/retrievers/self_query/tests/memory_self_query.int.test.ts

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[ykshev]

Unfortunately

const filter = { 'userId': { equals: userId } };
const retriever = vectorStore.asRetriever(undefined, filter);

this doesn't work :(

There are two methods which should get a filter argument and how I think should provide a similar output, but in first case filter doesn't work and it works in second:

await vectorStore.similaritySearch("hello world", 3, { userId: { equals: 1 } } )

await vectorStore.similaritySearchWithScore("hello world", 3, { userId: { equals: 1 } } )

I think it could be somehow related to my initial question

[aaditya25052002]

yeah, it works with only similaritySearchWithScore, btw do you have any idea how to apply filters for multiple fields?

[initcore0]

I've encountered the same problem. It appears that the example provided at https://js.langchain.com/docs/integrations/vectorstores/prisma suggests constructing a vector store using the static method withModel, which seems to completely ignore the filter parameter.

I managed to get the filter working by constructing the class instance directly and passing the filter like this:

const embeddings = new OpenAIEmbeddings()

export const newPrismaVectorStore = (userId: string) => {
    return new PrismaVectorStore(embeddings, {
        db: prisma,
        columns: {
            id: PrismaVectorStore.IdColumn,
            content: PrismaVectorStore.ContentColumn,
        },
        prisma: Prisma,
        tableName: 'Document',
        vectorColumnName: 'vector',
        filter: {
            user_id: { equals: userId },
        },
    })
}

Additionally, this approach doesn't seem to correctly map the table name/column name if you've used @Map(...) in your Prisma schema. So, in my example, I used user_id instead of userId.

I hope this helps

[Speuce]

Created an issue for this: #4521