Ask a network AI agent about OSPF and it answers from training data. Ask it what’s on your network right now and — if it’s built right — it goes and looks. But ask it about your design document, the vendor install guide for your platform, the customer standard that says which BGP communities mean what in your shop, and until this week NetClaw had nothing to reach for.
That gap is now closed by BISQUE — the Base of Ingested Semantic Queries for Unified Embeddings — NetClaw’s fully offline, free, local document knowledge base. Yes, it’s a backronym. No, we’re not sorry: the tastiest thing you can make from claws is a bisque.
Four kinds of knowing
The design started from a distinction that sounds academic and turned out to be the whole feature. An agent has four knowledge sources, and they must never blur:
- What it knows — parametric knowledge from training. Fine for "what is OSPF."
- What is true right now — live MCP calls (pyATS, NetBox, and friends). The only acceptable source for current network state.
- What it has experienced — its own Memory MCP: facts it learned, decisions, session history.
- What it has been given — and this is the new one: documents users deliberately hand it. Install guides, RFCs, vendor configuration guides, design docs.
BISQUE holds only category four. It is physically separate from Memory — separate directory, separate database, separate tools — and neither system ever writes into the other. We rewrote NetClaw’s SOUL and skills so it can articulate which source a question calls for, and the validation suite checks the routing logs for zero cross-source violations.
Retrieval is a tool, not a pipeline
Most RAG you’ve seen is a pipeline: every question gets embedded, top-k chunks get stuffed into the prompt, generation proceeds whether the chunks helped or not. BISQUE’s retrieval is agentic: rag_search is a tool the agent chooses to call — or not — and it can critique its own results and re-query, bounded by an explicit three-round budget so it can’t spiral. Under the hood it’s hybrid search — dense vectors in ChromaDB plus BM25 keyword scoring, fused and reranked by a local cross-encoder — because networking documents are full of exact tokens (router bgp 65001, part numbers, RFC numbers) that pure semantic search fumbles.
Two rules keep it honest. Every claim that comes from a retrieved document carries a citation — title, section, page, ingest date — and a claim it can’t cite, it doesn’t make from BISQUE. And when the corpus simply doesn’t contain the answer, NetClaw says so. Honest misses over hallucinated hits, 100% of the time.
Free and offline, on purpose
No cloud embedding APIs, no paid rerankers, no hosted vector database. Local embedding models, local reranking, everything under one directory. After the initial model download the whole thing works air-gapped — which matters to exactly the kind of network operators NetClaw is for. Ingestion takes PDFs, modern Office formats, HTML, and URLs; a LibreOffice fallback handles the legacy .doc/.xls fossils every enterprise still has. Users own the corpus: every document is listable, inspectable, deletable, re-indexable.
There’s one carefully fenced exception to "documents only": an opt-in snapshot tool can vectorize a chosen piece of live network output into a timestamped collection — for "compare this to last month" workflows — but it is never automatic, never in a heartbeat, and its age is always displayed. Live state stays the job of live tools.
First bowl of bisque: real numbers
The first real document went in this week — my own book, uploaded as a PDF attachment in Slack:
- 212 pages → 389 chunks, registered, embedded, and queryable in about 2 minutes 20 seconds end to end (the spec’s bar was 5 minutes for 100 pages)
- First cited answer sourced from the book seconds later, retrieval latency ~800 ms per hybrid search
- The agentic loop was visible in the retrieval log: two sub-queries came back at 0.96–0.998 confidence, one weak query got self-flagged as low confidence — the self-critique working on real traffic
- The honest-miss test: asked about vendor tech that isn’t in the book, every retrieved chunk scored below 0.03 and was flagged low-confidence — so the agent said "not in my knowledge base" instead of making something up
- Every ingest call was inspected and allowed by the DefenseClaw guardrail on the way in
The unplanned durability test
The same week BISQUE shipped, we upgraded the host straight through two Ubuntu releases to 26.04. The upgrade deleted the entire Python package layer out from under every MCP server — a genuinely rude afternoon — and when the dependencies were reinstalled, the knowledge base came back untouched: registry, vectors, retained originals, retrieval logs. State that lives in one well-defined directory survives what happens to the interpreter around it. Not a test we’d have volunteered for; nice to pass it anyway.
What’s next
The evaluation harness ships with a golden-set format but no fixture documents — the corpus should be real. So the next step is unglamorous and important: loading the bookshelf. RFCs, the vendor guides we actually use, the design docs that explain why the network looks the way it does. An agent that can check what it’s been given is only as good as what you give it.
NetClaw is open source: github.com/automateyournetwork/netclaw