Repository navigation
Ragmir: local, cited retrieval for LlamaIndex agents over MCP (open source) #24203
Replies: 1 comment
|
@jb-thery The noise floor mechanism effectively suppresses hallucinations for factual retrieval, but the "return nothing" strategy introduces a critical failure mode for reasoning tasks. When a query requires synthesizing information across multiple fragmented documents or correcting a logical inconsistency within the codebase, a hard refusal provides zero signal. The agent cannot leverage its reasoning capabilities to bridge gaps or flag contradictions because the retrieval layer treats ambiguity as absence of evidence. This forces the downstream model to hallucinate or loop rather than admit uncertainty about the data's completeness. A hybrid approach that returns the most relevant snippets with a confidence score, even if below the noise floor, allows the model to explicitly state "evidence is sparse" instead of fabricating an answer. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
I maintain Ragmir, an open-source (AGPL-3.0) retrieval layer that indexes a project's code, specs, PDFs, and Office files on the machine where they live. Each search returns passages cited to lines, PDF pages, slides, or spreadsheet cells, so an agent can show where every claim comes from.
LlamaIndex agents can use it through MCP with
llama-index-tools-mcp, without a Python package:Give
toolsto aFunctionAgentor any LlamaIndex agent.Two things that may be useful to people building agents here:
The full Python guide (MCP SDK, LangChain, LlamaIndex, Docker) is in docs/python.md. Feedback on the citation format and on the document types you need is welcome.
All reactions