qdrant_find not retrieving results

[outbound]

Running self-hosted qdrant on docker v1.13.6

I asked Claude to summarize the issue.

[LOG] Qdrant Integration Issue - Cannot retrieve but can store data
Environment: Claude AI Assistant (Windows 11) with Qdrant function integration

ISSUE SUMMARY:

• Successfully able to store data in <concept_name> collection using qdrant-store
• Unable to retrieve data with qdrant-find (returns error: 'document')
• Collection exists and is visible in Qdrant UI (1488 existing points, 4 segments)

COLLECTION DETAILS:

• Collection name: <concept_name>
• Vector configuration: fast-all-minilm-l6-v2 (dim 384, cosine distance)
• Contains 1488 points across 4 segments
• Collection status: green (active)

ATTEMPTED OPERATIONS:

1. SUCCESSFUL - Store operation:
   Function: qdrant-store
   Metadata provided: Valid metadata matching collection schema
   Information field: Text description of concept
   Result: "Remembered: [information text]"

2. FAILED - Retrieval operation:
   Function: qdrant-find
   Query: Various semantic search terms related to content known to exist
   Error: "Error executing tool qdrant-find: 'document'"

COLLECTION SCHEMA (from ingest script):

• Embedding fields: IM_Concept, ...
• Payload fields: ID, Pattern_Type...
• Vector embeddings created using: sentence-transformers/all-MiniLM-L6-v2

TROUBLESHOOTING ATTEMPTED:

• Varied query terminology and complexity
• Stored test points with properly formatted metadata
• Verified collection is accessible and properly indexed

[kacperlukawski]

@outbound Thanks for reporting the error! Could you please share a few points from your collection? I mainly mean the payload structure.

[outbound]

@kacperlukawski Here you go
https://docs.google.com/document/d/1lW4Vo_gRplytnB7v3GUSJUeQ1GLUo0B3X4C8VpGzdL8/edit?tab=t.0#heading=h.femsruuu1jg9

[outbound]

As a follow-up, I was also using Qdrant to store Claude's default memories. It can't retreive any of those either.
I setup ChromaDB and have had no issues with either collection.

[outbound]

@kacperlukawski Any update on this? Looks like there's another user with the same issue. #48

[Luttik]

I am having the exact same issue.

[kacperlukawski]

@outbound Looking at the screenshot, I can tell the collection was not created with our MCP server. The structure of the payload matters a lot, and the MCP server needs to have a payload with "document" and "metadata" fields, as shown below:

@Luttik Could you please also share an example of the point in your collection?

We have a plan to add another abstraction layer that will allow working with payloads of different structure, but for the time being, the payload structure has to be consistent with the one created by the MCP server.

[Luttik]

Hi, I think that will fix it for us as well, we're also not filling the vector-database throught the MCP. We want to be able to use the MCP to give plug-and-play read access to our Agents, rather than using it to store memory. I'll change the structure of what we publish and try again.