Basemind documents

basemind-documents — document RAG and web ingestion

basemind extracts 90+ file formats (PDF, Office, HTML, email, images via OCR) into a LanceDB vector store and answers meaning-based queries with cross-encoder reranking. Web pages scraped or crawled into the same store are searchable the same way. This is the surface for "find the passage about X", not "grep for the string X".

basemind first, open-the-file fallback. Prefer memory mode documents over opening PDFs/Office/HTML by hand, and the web tools over ad-hoc fetching. For source code use basemind-code-search instead — this skill is for prose and documents.

Requirements

• memory mode documents needs a build with --features documents (or full); the other memory modes need --features memory. Without them the tools dispatch but return an MCP error.
• Web ingestion (web modes scrape / crawl / map) needs --features crawl. When that feature is off these tools are not registered at all — they simply won't appear in the tool list.
• Documents must be scanned first: basemind scan with the documents feature extracts and embeds them into the machine-global cache (Linux ~/.local/share/basemind/, macOS ~/Library/Application Support/basemind/; override BASEMIND_DATA_HOME). See the basemind-scan skill.

Tool routing

Question	MCP tool	CLI
"Semantic search over PDFs/Office/HTML docs?"	memory { mode: "documents", query: "…" }	basemind memory documents "query"
"Narrow to docs mentioning an entity?"	memory { mode: "documents", query: "…", entity_category: "…" }	(MCP only)
"Narrow to docs with a keyword?"	memory { mode: "documents", query: "…", keywords_contains: "…" }	(MCP only)
"Filter by file type?"	memory { mode: "documents", query: "…", mime_type: "application/pdf" }	basemind memory documents "…" --mime-type application/pdf
"Pull a single URL into RAG?"	web { mode: "scrape", url: "…" } (robots-aware)	basemind web scrape <url>
"Ingest a docs site section?"	web { mode: "crawl", url: "…" }	basemind web crawl <seed-url>
"What URLs exist on this site?"	web { mode: "map", url: "…" }	basemind web map <url>
"Recall something the agent stored earlier?"	memory mode get, list, or search	basemind memory get "key" / list / search "q"
"Remember this for future sessions?"	memory { mode: "put", key, value }	basemind memory put "key" "value"

What a hit carries

memory mode documents returns chunk-level hits with path, chunk_idx, the matched text, byte span, vector distance, and — when enabled at scan time — a cross-encoder rerank_score in [0,1], the parent document's keywords and named entities (NER), and a document-level summary. Use entity_category / keywords_contains to constrain to documents whose parent carries a matching entity or keyword (AND-combined when both are set).

Examples

memory { mode: "documents", query: "how is the index schema versioned", limit: 5 }
→ docs/architecture.pdf#chunk3  rerank 0.91  "INDEX_SCHEMA_VER reads from RELEASE_MINOR…"
  README.md#chunk12             rerank 0.74  "…wipe-on-mismatch rebuilds from source…"

web { mode: "crawl", url: "https://docs.example.com/guide" }
→ ingested 24 pages under scope "web:docs.example.com"

memory { mode: "documents", query: "rate limiting", mime_type: "text/html" }
→ web:docs.example.com/limits#chunk1  rerank 0.88  "requests are capped at …"

Notes

• Crawled/scraped pages land in the documents table tagged with a scope of web:<host> (override in web mode scrape); memory mode documents searches every ingested document.
• robots.txt is honoured by default; only [crawl].respect_robots_txt = false in the repo-root basemind.toml (config-file-only) disables it. The crawler SSRF-blocks private/loopback hosts unless [crawl].allow_private_network = true.
• Memory is scoped by the normalised git origin URL, so clones of the same repo share stored entries and unrelated repos do not.
• Lists are capped (limit, default 100, max 1000); use next_cursor → cursor to page.

For code structure see basemind-code-search; for git history see basemind-git-history; for agent coordination see basemind-comms.