PGVector behavior is strange

[Eugene-M]

Hi!

I've noticed some peculiar behavior with the PGVector component. Ideally, PGVector should vectorize (or create an embedding of) the user's chat input, then find similar records in the database through a vector search. However, in practice, it seems to completely disregard the user's input, returning all records from the database in the search results regardless. Could you help me figure out what's going wrong?

[dosubot[bot]]

Hello @Eugene-M! I'm here to help you with any bugs, questions, or contributions you might have.

The issue you're experiencing with the PGVector component returning all records instead of finding similar ones is likely due to the search_query not being properly utilized. The search_documents method in the PGVector component checks if self.search_query is a non-empty string before performing a similarity search. If self.search_query is empty or not a valid string, the method defaults to returning all records.

To resolve this, ensure that the search_query is correctly set and is a non-empty string. This input is passed to the PGVector component through the MultilineInput named "search_query". Make sure that the chat input is being correctly transformed into a search query and passed to the PGVector component for the similarity search to be executed ^[1].

To continue talking to Dosu, mention @dosu.

[Eugene-M]

Unfortunately, your comment is not valid. I have already tried changing the code, substituting values manually, the search does not work properly. I even tried using other methods available in langchain_community.vectorstores.

[joaoguilhermeS]

Hey @Eugene-M, Did you Ingest your Data before trying to retrieve the information? I just tested it in the last Langflow release, and it seems to be working fine (V1.0.19). To allow your model to get only the important text parts for you, you need to split them using the Split Text Component, you can control each chunk length by the Chunk Size parameter, this way using the Embeedings parameters the component will check the most correlated chunks for your input.

I am attaching here an example Flow for your use case.
Simple Example PG Vector.json

[anjay-gupta]

@joaoguilhermeS - We are also facing the same issues , we have stored the json object and search text as josn and embedding used nomic-embed-text for data load ( preloaded ) , if we ask the context it return the all and if we ask the based on specific ID / attribute inside the Json ( which is already present in text and jsonb in vector db ) however this does not return the matching result and product id is already exists in the system.

Please let me know if we resolve this or any best approach to handle the semantic and search within the json attributes as well.

// Query generated the in chat when we specify the Product ID
SELECT langchain_pg_embedding.collection_id AS langchain_pg_embedding_collection_id, langchain_pg_embedding.embedding AS langchain_pg_embedding_embedding,
langchain_pg_embedding.document AS langchain_pg_embedding_document,
langchain_pg_embedding.cmetadata AS langchain_pg_embedding_cmetadata,
langchain_pg_embedding.custom_id AS langchain_pg_embedding_custom_id,
langchain_pg_embedding.uuid AS langchain_pg_embedding_uuid,
langchain_pg_embedding.embedding <=> '[-0.02035035379230976,-0.008137090131640434,-0.14020822942256927,-0.03820028156042099,0.06551223248243332, ...]' AS distance
FROM langchain_pg_embedding JOIN langchain_pg_collection ON langchain_pg_embedding.collection_id = langchain_pg_collection.uuid
WHERE langchain_pg_embedding.collection_id = '1d3628b0-6de2-4a9e-b64f-1000101001'::uuid::UUID
ORDER BY distance ASC
LIMIT 10