RAG

Skill: RAG Knowledge Base

Purpose: Give NetClaw a fully offline, user-curated document knowledge base — vendor guides, standards (RFC/IEEE/vendor), customer design documents, install guides — with agentic retrieval, mandatory citations, and opt-in point-in-time snapshots.

Overview

Users teach NetClaw by uploading documents (Slack attachment, HUD Knowledge panel, or URL). Documents are parsed, chunked structure-aware, embedded locally, and stored at ~/.openclaw/rag/. Retrieval is a tool NetClaw invokes on its own judgment — iteratively, with self-critique — never a fixed pipeline.

This is NOT memory. The knowledge base holds only what users deliberately put into it. NetClaw's own experience (facts, session summaries, decisions, entity graphs) lives in the Memory MCP (memory_* tools, ~/.openclaw/memory/). Neither store writes into the other.

MCP Tools

The Four Knowledge Sources (routing rules)

Route every question to the right source. Most questions need NO retrieval.

1. Parametric knowledge — timeless networking fundamentals (OSPF LSA types, BGP path selection). Answer directly. Do not search.
2. Memory MCP (memory_recall, memory_get_facts, memory_get_decisions) — NetClaw's own past sessions, learned facts, and decisions about THIS network. "What was that BGP issue last month?" goes here, never to rag_search.
3. RAG knowledge base (rag_search) — user-uploaded documents. Vendor procedures, customer standards, install steps. Check it BEFORE declaring ignorance on these topics.
4. Live MCP servers (pyATS, NetBox, etc.) — current network state. NEVER answer a live-state question from the RAG store. The only exception is an explicitly requested snapshot, whose age must always be shown.

When a question legitimately touches both Memory and the knowledge base, consult both — but attribute every part of the answer to its actual source. Memory recall is never presented as a document citation, and vice versa.

"Remember this document" → ingestion (RAG). "Remember that PE2 is in maintenance until Friday" → Memory (memory_record_fact).

Workflow: Slack Document Upload

When a user posts a file attachment in the Slack channel and asks NetClaw to learn it:

1. Download the attachment content.

2. Call the ingest tool with the base64-encoded file:

   python3 $MCP_CALL "python3 -u $RAG_MCP_SCRIPT" rag_ingest_base64 \
     '{"filename": "wlc-9800-upgrade-guide.pdf", "content_base64": "<b64>", "doc_type": "vendor"}'
   

3. Infer doc_type from the user's words ("this is our customer standard" → customer); default other.

4. Confirm in-thread with what was learned: title, doc_type, page count, chunk count, collection — plus the example_question from the response, so the user learns what they can now ask.

5. If the response is deduplicated: true, say the document was already indexed. If reindexed: true, say the prior version was replaced.

6. Errors (unsupported format, size cap, parse failure) are reported verbatim — never silently swallowed.

Supported formats: PDF, Markdown, HTML, TXT, DOCX, XLSX, PPTX, VSDX natively; legacy DOC/XLS/PPT/VSD when LibreOffice is installed.

Workflow: URL Ingestion (always preview first)

When a user asks to ingest a web page:

1. Preview: rag_ingest_url {"url": "...", "mode": "preview"} — returns the page title, the same-domain pages it links to (depth 1, capped at RAG_CRAWL_MAX_PAGES), and a scope_token. No ingestion happens.
2. Confirm scope with the user: "That page links to 6 same-domain pages. Ingest just the page, or all 7?" ALWAYS offer the single-page fallback. If the preview was truncated, say so.
3. Ingest:
   • Single page: rag_ingest_url {"url": "...", "mode": "ingest"}
   • Page + linked pages (only after explicit confirmation): rag_ingest_url {"url": "...", "mode": "ingest", "include_linked": true, "scope_token": "<from preview>"}
4. Confirm as with file ingestion. Each page records its own URL as source. PDFs served at URLs are handled automatically.

Never call mode="ingest" with include_linked=true without having shown the preview and obtained the user's confirmation — the server rejects a missing/stale scope_token.

The Agentic Retrieval Protocol

Retrieval is a tool you wield, not a pipeline you sit inside. For every question:

1. Route first (Adaptive RAG)

Decide whether to retrieve AT ALL using the four-source rules above. Most questions need no retrieval:

2. Rewrite and decompose

Rewrite conversational phrasing into retrieval-friendly queries ("how do I get the new code on the WLC" → "WLC software upgrade procedure install activate commit"). Decompose multi-part questions into independent sub-queries, retrieved separately with scoped filters, then synthesized. Give each sub-query a sub_query_id and pass round on every rag_search call so the budget is auditable.

3. Critique after every retrieval (Self-RAG)

Grade the returned chunks: do they actually answer the question?

• Yes → answer, citing every claim.
• Partially / no → rewrite the query (different terms, add filters) and re-retrieve.
• low_confidence: true on results → treat as signal, never as answer material.

4. Respect the iteration budget — 3 rounds per sub-query

Each sub-query gets at most 3 retrieval rounds (initial + 2 refinements; RAG_MAX_ROUNDS). Stop condition: when a sub-query's budget is exhausted, stop retrieving for it and state plainly what you found (cited) and what remains unanswered. Do not loop. Do not pad.

5. Honest miss

If the corpus doesn't cover the topic — empty results, corpus_empty: true, or only low-confidence chunks after refinement — say so:

"The knowledge base doesn't cover Nexus 9300 upgrades. I can ingest a document if you have one — drop it in this channel."

NEVER answer from irrelevant or low-confidence chunks. NEVER present a guess as knowledge-base content. A fabricated answer is worse than an honest gap.

Citation Rules (non-negotiable)

Every claim derived from retrieved content carries a citation:

[WLC 9800 Upgrade Guide §4.2, p.31 — ingested 2026-07-01]

Use the citation field returned by rag_search verbatim. A claim you cannot attribute to a specific retrieved chunk is NOT presented as coming from the knowledge base. Chunk IDs stay in the retrieval log — never show them to users.

When synthesizing across multiple documents, every combined claim must be supportable by at least one cited chunk. Do not blend unrelated chunks into a claim none of them supports — that is synthesis hallucination.

Snapshots — the ONLY sanctioned RAG use of live data

rag_snapshot exists for one purpose: explicit point-in-time comparison ("snapshot the BGP tables so we can compare next month"). It is NEVER a substitute for a live MCP query.

Absolute prohibition: NEVER invoke rag_snapshot automatically, from a heartbeat, on a schedule, or as a side effect of another task. It runs only on an explicit human request, and only after you confirm scope.

Workflow:

1. User explicitly asks to snapshot live output.
2. Confirm scope before executing (HIIL gate): which devices, which commands, what label. "Scope: PE1, PE2 — show ip bgp via pyATS, into snapshot_core-bgp_<timestamp>. Confirm?"
3. Collect the output via the existing MCP tools (pyATS etc.).
4. Call rag_snapshot {"label": "core-bgp", "content": "<output>", "source_description": "core router BGP tables", "devices": ["PE1","PE2"], "commands": ["show ip bgp"]}.
5. Report what was stored including the per-type redaction counts from the response — secrets (passwords, SNMP communities, auth keys, pre-shared keys) are scrubbed before vectorization, and a zero count is reported explicitly.

When ANY answer later uses snapshot data:

• Lead with the staleness notice: "From snapshot core-bgp — captured 2026-07-16 14:02 UTC — 31 days ago (live state is available via MCP tools): …" Use the age_human/staleness_notice fields verbatim.
• Remind the user that live state is one MCP call away.
• Snapshots older than RAG_SNAPSHOT_WARN_DAYS (default 90) are flagged stale — surface the flag. They are never auto-deleted; offer deletion instead.

Environment Variables

Example Usage

User: (attaches customer-wlan-standard.docx) learn this, it's our customer standard
NetClaw: Learned "Customer WLAN Standard" (customer, 24 pages, 61 chunks, documents).
         Try asking: "What does our standard 'Customer WLAN Standard' say about maintenance windows?"