Agent Search and Retrieval

Agent search is routed by source of truth: qmd for Kevin's compiled wiki, Graphify for repo topology, docs lookup for API facts, skills.sh for installable procedures, and web/social search only for current external evidence. Source: User request, 2026-06-27; Capability Routing Map

When To Use This

Use this page when an agent needs to find knowledge, docs, tools, skills, or current web evidence before acting. Do not use browser automation as a substitute for search. Source: The Brain-Agent Loop, 2026-06-27

Decision Matrix

Core Model

Retrieval quality depends on matching the question to the right corpus. qmd answers "what has Kevin already compiled?" Graphify answers "what is connected in this codebase?" Docs lookup answers "what does this API currently say?" Skills discovery answers "is there a procedure I can install?" Web search answers "what changed outside the wiki?" Source: Capability Routing Map, 2026-06-27

Scale changes the retrieval contract. A flat markdown wiki should remain the editable source of truth, but once the corpus is large enough that ad hoc grep returns noisy context, search becomes a first-class subsystem with indexes, reranking, source provenance, and explicit routing. The GBrain comparison bookmark frames this as the "LLM-Wiki hits a wall" failure mode; Kevin's answer is not to abandon files, but to put qmd or a hybrid retrieval layer between the files and the agent. Source: X/@JoshuaIPark, 2026-04-10; Source: Knowledge Brain System

Tool And Skill Routing

Start brain-first with qmd unless the user explicitly asks for live/current external information or codebase topology. Use rg for exact local text. Use Graphify for structural repo questions where a focused subgraph, path, explanation, or affected-node set can reduce raw file reading. Use docs tools for versioned APIs. Use skills.sh only for procedure discovery, not general facts. Source: Skill Resolver, 2026-06-27

Failure Modes

• Searching the web before the wiki for Kevin-specific knowledge.
• Treating skills.sh results as factual documentation.
• Using qmd for facts that must be current outside the wiki.
• Browsing pages manually when a docs/search interface is available.
• Treating screenshot retrieval as magic when auth, collapsed sections, or late-loading content were never rendered.

Evidence And Timeline

Good retrieval leaves source evidence: the page read, qmd query, doc URL, skill slug, or web source date. Reusable discoveries should be compiled into wiki pages or skills. Source: The Brain-Agent Loop; No One-Off Work

---

Timeline

• 2026-07-02 | Added Graphify as the repo-topology retrieval route while keeping qmd as the compiled-wiki retrieval layer. Source: User request, 2026-07-02

• 2026-07-04 | Added the GBrain/LLM-Wiki scale boundary: markdown stays canonical, while qmd or hybrid search becomes mandatory when grep-style retrieval stops being precise enough. Source: X/@JoshuaIPark, 2026-04-10

• 2026-06-30 | Added screenshot/visual retrieval routing after PixelRAG review. Use it when rendered layout carries the evidence, but record capture state and browser actions as part of provenance. Source: X bookmark artifact audit, 2026-06-30

• 2026-06-27 | Canonical cluster page created to separate qmd, docs lookup, skills.sh discovery, and live web/social search. Source: User request, 2026-06-27