Build a knowledge graph from a local directory of docs and/or code files.

Docs (md/markdown/mdx/txt/wiki/rst/adoc) are indexed as-is with edges from markdown links and [[wikilinks]]; code files contribute import/include edges and searchable symbol headers. Returns site_id plus a graph-quality verdict: READY (traversable edges found) or NOT-READY (no links/imports; use search).

relations selects the derived relation dimensions, chosen per node: "auto" (keyless: mentions_symbol/similar_to/sibling_of/supersedes/ shares_heading/shares_term/duplicate_of/coupled_with/co_cited_with), "llm" (auto + an LLM router plus entity/embedding similarity), "all", "off", or an explicit comma list. Omit it to use $SITE_KG_RELATIONS, or "auto" when that is unset. Content-derived edges also run on corpora with no cross-references; the verdict still reports NOT-READY for those.

Crawl a website, build its knowledge graph, and register it as a queryable site.

Returns site_id for later tools plus a graph-quality verdict: READY (traversable edges found) or NOT-READY (tree without links; graph traversal useless).

relations selects the derived relation dimensions, chosen per node: "auto" (keyless dimensions), "llm" (LLM router + entity/embedding similarity), "all", "off", or an explicit comma list. Omit it to use $SITE_KG_RELATIONS, or "auto" when that is unset.

dev is the development/testing mode for sites you own or are authorized to test: it ignores robots.txt and politeness delays so a deep crawl of an authorized target finishes fast. It forces respect_robots=False.

List ingested sites with their verdicts and graph sizes.

Graph health for one site: doc/edge counts, edge types, verdict, top hubs.

Full-text search over the site's pages (inverted index, idf-weighted).

Full markdown body of one page, with title and source URL.

Subgraph around a page up to depth hops (capped at 2). Use after search to expand context.

Render the graph as a self-contained HTML viewer (human inspection path). Archive mode returns the file path; mode="live" replays the stored graph over a loopback SSE server and returns {live_url, run_id, last_seq}. AI-facing queries stay on search/ask/neighbors.

Render the graph construction process as a self-contained WebGPU viewer (human inspection path). Replays the build: nodes are born in corpus order onto chapter rings, then evidence edges (ref/imports/contains), then one flying wave per derived relation dimension -- every edge grows from its source node and binds to its target with a pulse, while a HUD counts nodes, edges and per-dimension edges as they land. Returns the file path; AI-facing queries stay on search/neighbors/ask. Needs a WebGPU-capable browser.

LLM Q&A grounded in the site's pages: keyword retrieval + graph-neighborhood expansion, then embedding rerank and a cited answer. Requires SITE_KG_LLM_BASE and SITE_KG_LLM_KEY in the server environment.