Knowledge Brain

A real, small, honest search tool -- not an AI memory system.

Every video Project KAI produces leaves behind real evidence: a CEO report, a compliance report, a fact-check note, a quality score. The Knowledge Brain is the tool that searches across that evidence and the project's own engineering reports. As of the latest data sync (August 11, 2026), it indexes107 real reports. It is not semantic search, it does not maintain a persisted index, and it is notwired into any automatic process today.

What it indexes

107 real reports

Root-level reports: 62 -- engineering, launch, and governance .md documents at the repository root.
Per-video reports: 45 -- ceo_report.md, compliance_report.json, fact_check.json, and quality_score.json for each produced video.

How it searches

Keyword overlap, not meaning

Word-overlap keyword search (lowercase alphanumeric tokens, ranked by intersection count). Not semantic search -- no embeddings, no vector index.

Example, straight from the tool's own source comment: a query like"have we ever used this thumbnail?" will find reports containing the word "thumbnail" -- not reports that discuss the same idea using different words. That's the honest limitation of keyword search, stated plainly rather than papered over.

What it doesn't do

No index, no automation

No persisted index -- every search re-reads the real files fresh. Small enough corpus that this can never go stale, at the cost of taking slightly longer per search.
Not automated -- a repository-wide check found zero callers outside its own package. It's invoked directly by an engineer, not triggered by production, publishing, or analytics.
Read-only -- it never writes, moves, or deletes anything it indexes.

How evidence actually moves

From production output to a manual search.

Active

Production output

↓

Active

Real reports written per video (CEO report, compliance report, fact-check, quality score)

↓Manual invocation only

Manual

Knowledge Brain search (knowledge_brain/indexer.py)

↓

Manual

Human reads the results and decides

This is not an automatic pipeline. Nothing in the real production or publishing code calls the Knowledge Brain today -- an engineer runs a search directly when they want to check something.

Current

Searches all 107 real reports on demand
Ranks results by keyword-overlap count
Read-only -- never modifies what it indexes
Callable directly as a Python function today

Planned

Upgrade to semantic search -- gated on the report corpus growing large enough that keyword search's limitation starts actually costing time, per the project's own real roadmap
Automatic wiring into production runs, so lessons surface without a manual search

Technical Details

knowledge_brain/indexer.py -- all_reports(), search()
Report categories: root_report, topic_report:<video-id>
Search: lowercase alphanumeric tokenization, ranked by set-intersection size between query and content
KAI OS knowledge_brain/indexer.py (all_reports/search)