knowledge-graph - SKILL.md Agent Skill

name: knowledge-graph description: Search and query the ForgeRAG engineering knowledge graph. Ask questions about materials, welding, standards, formulas, reference tables, and specifications — get answers with page citations from ASM handbooks, ASME codes, NFPA, IEEE, and other engineering references. version: 1.2.0 author: system tools: - ask_engineering_question - find_relevant_chunks - search_engineering_docs - smart_search - get_forgerag_status - query_knowledge_graph - explore_entity - list_knowledge_collections dependencies: []

Engineering Knowledge Graph (ForgeRAG)

Search engineering handbooks, standards, and specifications. The knowledge graph contains materials, processes, standards, equipment, formulas, reference tables, and their relationships extracted from ingested PDF documents — with paragraph-level structural chunking, LLM-generated chunk summaries, RRF hybrid retrieval (BGE-M3 dense + BM25 + cross-encoder reranker), fuzzy entity matching, OCR typo tolerance, community weighting, and a circuit breaker for reliability.

When to Use

Tool	Use when…
`ask_engineering_question`	The user wants a synthesized answer with citations. Uses RRF hybrid + VLM reading of page images. Primary tool.
`find_relevant_chunks`	You need precise evidence to quote (a specific paragraph, table, or equation). Returns raw chunk text + summary without the VLM synthesis step. Faster than ask_engineering_question.
`search_engineering_docs` (mode="keyword")	Looking up a specific code, alloy designation, clause number, or standard (C12000, QW-451.1, NFPA 70, SEMI S2). Supports `fuzzy: true` for OCR typo tolerance.
`search_engineering_docs` (mode="visual")	Finding pages by visual similarity — charts, tables, diagrams, schematics. Uses Nemotron ColEmbed.
`smart_search`	General-purpose entry point when unsure which mode is best. Auto-detects: codes/designations -> keyword, questions -> answer, else -> hybrid/RRF.
`get_forgerag_status`	Check ForgeRAG capabilities, live stats (documents, pages, entities, communities), and service health before searching.
`query_knowledge_graph`	How entities relate to each other — what standards govern a material, what materials are compatible with a process, which standards cross-reference.
`explore_entity`	Everything connected to one specific entity (N-hop neighborhood).
`list_knowledge_collections`	Discovery — what engineering databases are available.

Tool selection patterns

"What does ASTM A36 require for…" → ask_engineering_question (narrative answer)
"Quote the exact text of QW-451.1" → find_relevant_chunks (precise quotation)
"Find the tap drill table" → find_relevant_chunks with chunk_type="table"
"What formula is used for beam deflection?" → find_relevant_chunks with chunk_type="equation", then ask_engineering_question if more context needed
"Show me the torque chart" → search_engineering_docs with mode="visual" (page images)
"Tell me about Alloy 625" → smart_search (auto-routes to the best strategy)
"Find C1200O in scanned docs" → search_engineering_docs with mode="keyword" and fuzzy: true (OCR typo tolerance)
"What collections and entities are available?" → get_forgerag_status (live stats and capabilities)

Important

ask_engineering_question is the primary synthesis tool — RRF hybrid retrieval + VLM reading pages + graph context → answer with [Page N] citations
find_relevant_chunks returns paragraph/table-precise results with LLM summaries — use it for precise quoting or when you want to inspect the raw evidence before synthesizing
For specific alloy codes (C12000, A36) or clause IDs (QW-451.1, NFPA 70-15A §7.2.3), keyword search often nails exact matches faster than the hybrid retriever
The graph stores Materials, Processes, Standards, Equipment, Formulas, and RefTables — use explore_entity to discover cross-references and relationships
Standards have BOTH a short code (e.g., "SEMI S2") and a title (e.g., "Environmental, Health, and Safety Guideline for Semiconductor Manufacturing Equipment") — queries resolve either form
Collections organize documents by domain (asm_references, mechanical_design, firearms, etc.) — specify a collection to narrow searches
Page-level topic_tags classify content topically (tap-drill-chart, fastener-torque, conductor-ampacity, etc.) — helpful for filtering
Answers include page numbers — cite them when reporting findings
The system reads actual page images via a VLM, so it can interpret tables, charts, and diagrams that text extraction misses