YOUR AI PLATE HANDBOOK
Documents & hybrid retrieval
Understand what belongs in the knowledge graph and what belongs in document search.
Desktop source reviewed October 5, 2026
Two complementary knowledge sources
| Tier | Best suited to | Retrieval behavior |
|---|---|---|
| OKF knowledge graph | Reviewed rules, specifications, definitions, and relationships | Matches structured nodes and traverses related links. Canonical context is formatted before supplementary document chunks. |
| Vector RAG | Manuals, notes, and source documents | Retrieves embedded text chunks by similarity using the configured embedding model. |
The router supports hybrid, okf_only, and rag_only modes at the implementation level. These are router options, not three promised UI buttons. Retrieved context helps ground an answer; it does not guarantee that the answer is correct.
Add and query a document
- Configure an embedding provider and model in Models & Reasoning.
- Open Knowledge Base, or use the chat attachment control to add a supported text-based file.
- Wait for ingestion, then check the document and chunk counts.
- Ask a specific question and compare the answer with the original passage.
The parser handles PDFs, Markdown, text, code, JSON, CSV, and TSV. Text extraction is required; the parser does not implement OCR for scanned pages.
When a document is not found
| Symptom | Check |
|---|---|
| Chat works, ingestion fails | Embedding configuration and the selected endpoint’s embedding support. |
| Empty or garbled PDF text | Whether the source PDF contains extractable text. |
| Wrong or incomplete answer | Retrieved passages, query specificity, and the source document; verify numerical results separately. |
| OKF is available but RAG is empty | OKF lexical/graph search and vector ingestion are separate paths. |
Implementation sources
Paths in your AI Plate source checkout:
core/document-parser.tscore/vector-store.tscore/embedder.tscore/hybrid-knowledge-router.tsui/index.html