Skip to content

Grounded Reasoning

The grounded reasoning subsystem is the closed-corpus reasoning path behind sb grounded.

Purpose

  • answer questions against local corpus evidence
  • run grounded evals
  • generate synthetic grounded QA data
  • inspect stored reasoning trajectories and reporting

Current CLI Surface

sb grounded answer "What is SecondBrain?"
sb grounded eval
sb grounded synth "machine learning" "vector databases"
sb grounded replay
sb grounded report

Use sb grounded --help and subcommand help for the exact flags.

Runtime Model

The subsystem follows a multi-step reasoning loop:

think -> retrieve -> observe -> think -> answer

Core package areas:

  • brain/grounded/harness/
  • brain/grounded/retrieval/
  • brain/grounded/plugins/
  • brain/grounded/evals/
  • brain/grounded/store/

Typical Workflow

Answer a question

sb grounded answer "What is in the local corpus?"

Evaluate the subsystem

sb grounded eval --json

Generate synthetic QA data

sb grounded synth "memory systems" "retrieval" "reasoning"

Inspect prior runs

sb grounded replay
sb grounded report