Progressive-disclosure query interface for the Obsidian LLM wiki. Returns the cheapest useful answer first: a 1-5 sentence RLM summary. The caller can then request bullets, then the full wiki node — expanding context only as needed.

Progressive Disclosure Levels

Level	Content	Cost
`summary`	1-5 sentence distilled answer	~50 tokens
`bullets`	6-10 key idea bullets	~150 tokens
`full`	Complete wiki node + wikilinks	~800 tokens
`raw`	Original source file content	variable

Usage

Quick summary (default)

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root "authentication flow"

Bullet-level detail

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root "authentication flow" --level bullets

Full wiki node

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root "authentication flow" --level full

File result back into wiki (Karpathy's "outputs always add back" loop)

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root "rlm design" \
    --level full --save-as my-rlm-research

Use specific vector DB profile for Phase 2

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root "attention mechanism" \
    --vdb-profile research

Use shared .agent/learning/ RLM cache

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root "auth flow" \
    --rlm-cache-dir /path/to/project/.agent/learning/rlm_wiki_cache

List all indexed concepts

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root --list

JSON output (for programmatic use / agent pipelines)

python ./scripts/query_wiki.py --wiki-root /path/to/wiki-root "api design" --json

Search Strategy (3-Phase)

Phase 1 — Slug/token match (O(1), always runs):

Exact concept slug match
Slug is a prefix/substring of a concept name
Shared word-token overlap (e.g. "auth" matches "authentication-flow")

Phase 2 — Vector DB semantic search (O(log N), requires vector-db installed):

Calls vector-db plugin's query.py as a subprocess
Resolves vector DB config from .agent/learning/vector_profiles.json
Default profile: wiki (override with --vdb-profile)
Maps semantic results back to concept slugs via meta/agent-memory.json
Gracefully skipped if vector-db is not installed

Phase 3 — Full-text keyword scan (O(N), always available):

Grep-style scan of wiki/*.md content as final fallback

--save-as: Filing Results Back Into the Wiki

Karpathy's key insight: "I end up filing the outputs back into the wiki to enhance it."

The --save-as flag writes the query result as a new wiki node:

wiki/{concept-slug}.md  ← new concept page derived from the query result

The saved node includes:

YAML frontmatter with query_derived: true and derived_from attribution
Original content at the requested disclosure level
## See Also link back to the source concept

This means every query session can grow the wiki, not just read from it.

When to Use

Any time you need fast context about a concept in the wiki
Before reading a full raw source file (use summary first)
When /wiki-ingest has been run and nodes are populated
As a pre-flight check before expensive agent operations

Related Scripts

query_wiki.py — progressive-disclosure query engine
raw_manifest.py — WikiSourceConfig for path resolution
audit.py — reports missing or stale nodes

obsidian-query-agent

Dependencies

Obsidian Query Agent

Purpose