Add Retrieve API/MCP

Use Case

I know the original premise of this project was to move beyond traditional RAG, but the UI and overall system are far ahead of most other tools in this space as of 2026-09-16.

It would be useful to use this system as a single home for documents while allowing AI agents to access that information in two ways:

• Graph-based recall/reasoning for higher-level knowledge and relationships.
• Traditional embedding-based retrieval when the original document text or high-detail technical information is needed.

This would be especially useful for things like SKILL.md files, technical documentation, configuration references, API documentation, and other material where verbatim retrieval is important.

Problem Statement

Graph based retrieval can have too low of a fact/detail resolution for technical documents.

During document processing, small but important details may not become graph nodes or facts, even though those details still exist in the source document. This is particularly noticeable with technical documentation where a single sentence may be important.

The reflect endpoint can help with deeper retrieval, but it is too slow to use as the default retrieval mechanism for frequent agent/tool calls.

Another RAG/vector database currently needs to be maintained alongside this system just to provide direct document retrieval.

How This Feature Would Help

Adding basic document-level semantic retrieval would allow this system to become a single source of truth for both graph knowledge and traditional RAG.

An AI agent could use graph retrieval when it needs summarized knowledge, relationships, or reasoning, and direct retrieval when it needs exact technical details from the underlying documents.

This would eliminate the need to ingest and maintain the same documents in a separate vector database and would make the system particularly attractive as a general knowledge backend for AI agents.

Proposed Solution

Add a quietly exposed background API/MCP tool called "retrieve".

This endpoint would perform ordinary embedding/vector similarity search directly against the original document chunks rather than against generated graph nodes or extracted facts.

Alternatives Considered

I want it so bad that I'm debating if I can make some sort of fork of this repo that AI auto updates it when you do releases. I have tried every other RAG/graph tool I can find and this was the easiest and prettiest to setup - it's 1 feature away from being an ultra tool that make is a no-brainer to pick vs the others like GraphRAG, Cognee, LightRAG.

Priority

Important - affects my workflow

Additional Context

No response

Checklist

• I would be willing to contribute this feature