Retrieve queries based on metadata in a Qdrant database

[Mohana-Murugan]

Hi,

I am working on a RAG-based LLM agent. Here, I am storing multiple user details and document details in a single collection, organized based on user_id and agent_id. I am writing a search approach to retrieve information from the database based on these IDs. However, I am facing an issue where the results return information even if the specified information is not in the document. Additionally, how can I isolate data when the same files are stored for different users?

''' query = "Bullying"
query_vector = get_embedding(query)

Search with a filter to match only documents containing "Qdrant"

results = client.search(
collection_name='test',
query_vector=query_vector,
limit=1,
query_filter={
"must": [
{"key": "user_id", "match": {"value": "K5617"}},
{"key": "agent_id", "match": {"value": "5307"}},
]
}
)

Check if results contain any data

if results:
for result in results:
print(f"Document ID: {result.id}, Score: {result.score}")
print(f"Text: {result.payload['text']}\n")
else:
print("Information not available")'''

[timvisee]

However, I am facing an issue where the results return information even if the specified information is not in the document.

I'm not exactly sure what your approach is like, because your example does not make that very clear.

If you're using payload values to isolate searches, that should already be sufficient. If it is problematic somehow I expect there to be a bug in the implementation.

If you're embedding your query and are searching using vector similarity, results can be approximate. For example, if there are no close matches to your search query, it'll simply return the closest match it can find even if it's completely unrelated material.

Additionally, how can I isolate data when the same files are stored for different users?

You can isolate data by specifying a payload value that is unique across users. Please make sure you configure a payload index for this, as it'll help to keep searches fast and to keep good search accuracy.

[Mohana-Murugan]

Yes, I am not able to get the right answer based on a filter.

[Mohana-Murugan]

So, we have planned to move Pinecone vector DB because of these issues.

[timvisee]

If what I describe above is what's happening - Pinecone will not fix this for you and give you exactly the same behavior.

Can you reduce the problem you're seeing to an example snippet of about 10 lines of code? Then I'd be happy to help you out with this.

For example, if we take the following snippet from your original comment, what response does Qdrant give you?

results = client.search(
collection_name='test',
query_vector=query_vector,
limit=1,
query_filter={
"must": [
{"key": "user_id", "match": {"value": "K5617"}},
{"key": "agent_id", "match": {"value": "5307"}},
]
}
)

Up until now it's still not clear to me what the exact issue is. Because I'm afraid that 'I am not able to get the right answer based on a filter' is rather vague.

[Mohana-Murugan]

My exact issues are as follows

• I have stored the RAG PDF under: user_id: K5617, Agent_id: 5307, Company ID: 101.
• I have stored the HR Policies under: user_id: K5618, Agent_id: 5308, Company ID: 102

It works fine:

When IDs do not match:

It provides the following wrong answer

[Mohana-Murugan]

I am using the same embedding model, 'BAAI/bge-small-en-v1.5,' for storing and retrieving. However, I cannot identify the reason why I am getting these results.

When I pass this to the LLM, the model provides answers based on its pre-trained documents instead of retrieving information from Qdrant using a filter.

Without any filter, Qdrant works fine for us, and there is no doubt about that. (I have attached the results of with and without LLM and filter condition above)

If you provide a solution for this issue that would be helpful for us. Million thanks in advance..!