· 中文 · Documentation
Technical documents · Engineering entities · Hybrid retrieval · Claim verification · MCP · Native STEP/STL
RagKernel is a verifiable engineering knowledge engine for building evidence-grounded systems over documents, CAD models, and equipment data. Additional engineering formats are planned behind the same ingestion contract.
Ask in natural language → RagKernel retrieves evidence first, then answers with a per-claim citation back to the source document and section. If it isn't in the corpus, it says so.
The installer touches your database, API keys, and documents. For security-sensitive environments, review the script before execution:
curl -fsSL https://raw.githubusercontent.com/v0id-byte/ragkernel/main/install.sh -o install.sh
less install.sh
sh install.shOr pipe it directly:
curl -fsSL https://raw.githubusercontent.com/v0id-byte/ragkernel/main/install.sh | shIt provisions the runtime (uv / Python 3.12 / dependencies), then walks you through LLM provider → local models → admin account → MCP integration. Run ragkernel doctor any time to self-check.
Prefer to do it by hand:
uv sync # add --extra cad for native STEP/STL
uv run ragkernel setup # provider / admin / models / MCP token
uv run ragkernel models # one-time local embedding + reranker download (~2GB)
uv run ragkernel serve # open http://127.0.0.1:8360Drop manuals, tickets, or CAD files into the web UI → they are indexed automatically → ask a question → get an answer with traceable citations.
Installer flags, Docker, and non-interactive/CI installs → docs/installation.md.
While developing embedded systems and designing PCBs, I repeatedly searched through hundreds of pages of datasheets and reference manuals. Existing RAG pipelines could retrieve relevant text, but they flattened engineering documents into chunks — losing engineering context and the link between an answer and its original evidence. RagKernel preserves engineering structure, evidence, provenance, and geometry, so humans and AI agents retrieve knowledge that is traceable and verifiable.
- Engineering document ingestion — PDF (Docling + RapidOCR, real page numbers), DOCX / PPTX / HTML, Markdown / TXT, CSV / XLSX tickets.
- Native CAD ingestion (STEP / STL) — assembly trees, exact B-rep geometry, mesh validity; optional
[cad]extra. See docs/cad.md. - Element-aware chunking — spec / fault-code / pinout tables split one row per chunk, procedures kept whole, engineering dimensions preserved, each chunk carrying structured metadata.
- Hybrid retrieval — BM25 + vector fused by RRF, reranked by a local cross-encoder, with exact metadata filtering (
search_by_field). - Traceable citations — every result carries stable document and chunk identifiers, plus source page numbers when the format and parser provide them.
- Evidence-backed claim checking —
verify_engineering_claimreturns supported / contradicted / unsupported with real page citations. - Explicit provenance — CAD measurements distinguish
brep_computed,mesh_computed, andfile_declared; invalid meshes report an invalid state instead of a misleading volume. - MCP server — read-only retrieval exposed to agents (stdio + HTTP, token auth, tiered rate limiting) alongside CLI and Web.
- Operable by default — one-line deployment, guided
ragkernel setup, read-onlyragkernel doctorwith JSON output for monitoring.
Full capability matrix and supported formats → docs/capabilities.md.
Full documentation lives in docs/.
| Doc | What's in it |
|---|---|
| Installation | installer flags, manual install, Docker, platform requirements |
| Configuration | provider setup, config precedence, ragkernel setup |
| Capabilities | format matrix, retrieval, guarantees, current limits |
| CLI reference | all commands + verification and eval scripts |
| Architecture | engine layers, code map, document lifecycle |
| Native CAD | STEP/STL formats, exact vs approximate geometry, CAD tools |
| Diagnostics | ragkernel doctor contract, exit codes, JSON schema |
| Web UI | ingestion, dashboard, admin console |
| Design principles | principles and philosophy |
RagKernel parses text, tables, pin definitions, dimension callouts, and scanned content in technical PDFs with real page citations, and reads verifiable STEP/STL geometry natively. It does not read DWG / SLDPRT / Parasolid natively, rebuild parametric feature trees, or perform full GD&T and hole recognition (a cylindrical face is not a confirmed hole). Page citations follow the parser's element provenance — rows of a table spanning multiple pages may currently cite the table's starting page. Single-tenant MVP; multi-tenancy and RBAC are future work.
Licensed under the Business Source License 1.1 — source-available. Free for personal, educational, research, and internal business use; each released version converts to Apache 2.0 on 2029-07-21. See LICENSE for the full terms. Commercial hosted or managed services offered to third parties require a separate commercial license.
Created and maintained by v0id-byte. Copyright © 2026 Liuhaoran Qin.
