Skip to content

FM: Router model — small local LLM for structural decisions #28

Open

Description

@diskerror

A local 3–8B model (same endpoint as [inference.memory]) that makes structural payload decisions before each turn:

Decision Output
Recipe selection recipe name
RAG query rewrite search string
Subject-switch detection continue / switch / branch
Summary-merge feasibility merge_ok / start_new
Per-source budget nudges widen/narrow list
Capture-worthiness store / skip

Open questions (carry as comments on this issue):

  • Blocking vs. parallel/speculative execution?
  • Cache recipe decisions per session?
  • JSON / grammar-constrained decoding for safe outputs?

Fallback: if no router configured → deterministic rules, recipe = default.

Activity

  1. added
    backendServer / API work
    fading-memoryFading memory / chat manager design
    chatRagger Chat (Role 2) / proxy (Role 3)
    on Apr 20, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    backendServer / API workchatRagger Chat (Role 2) / proxy (Role 3)fading-memoryFading memory / chat manager design

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions