Mimiron AI

What's next

RAG and the agent are live and measured. Where each framework stands, and what is next.

used narrowly

LangChain

Its text splitters cut the documents for two of the three chunkers. They replaced my own splitters after I measured both on the same files: identical chunks for recursive, and 8 tiny fragments instead of 43 for fixed. Search, prompting, generation and tracing are written by hand, on purpose.

live

MCP

The agent reaches the eval database through Neon's MCP server. A guard keeps only the read-only tools, rejects risky SQL, and redacts secrets from tool output. The project's own search is handed to the agent as a local tool rather than over MCP, because nothing else needs to call it.

live

LangGraph

The agent on /agent is a LangGraph loop: model and tools under a measured step limit, then one self-check that grades the answer against the evidence the tools returned. Every run streams its trace to the page.

The agent work, in order

What was planned, and where each step stands.

  1. Search and the latest eval results exposed to the agent as tools — done
  2. Database access through Neon's MCP server, behind an SQL guard and secret redaction — done
  3. A LangGraph loop with a measured step limit and a self-check on every answer — live on /agent as the owner's demo
  4. A hand-written loop, and the published comparison against LangGraph — not started