Glossary

Source — an uploaded document (PDF, markdown, audio, video, image, etc.). Disambiguate from "source code" when both might appear.

Loader — a plugin that converts an uploaded file into normalizable text. Eight built-in loaders cover PDF, markdown, plain text, audio, video, image, archives, and structured data.

Normalization — preprocessing that fixes encoding, removes structural noise, and produces clean text from a loaded source.

Chunk (small chunk) — a ~225-token segment used as the unit of retrieval. Chunks have stable IDs derived from content hashes.

Group (hierarchical group) — a ~900-token aggregation of small chunks used as the unit of entity extraction. Groups summarize and provide context across multiple chunks.

Indexing — the stage that produces embeddings and full-text indexes from chunks. Embeddings go to sqlite-vec; full-text goes to FTS5; both live in app.db.

Entity extraction — the stage that asks an LLM to find nodes and edges in groups. Constrained by domain templates.

Domain — a .jsonld file declaring node templates, edge templates, and prompts for a specific subject area (e.g., research, recipes, philosophy). Sixteen built-in domains; users can add custom domains.

Node template / Edge template / Graph template — reusable schema fragments. Distinct: node templates describe entities, edge templates describe relationships, graph templates compose the two for a domain.

Deduplication — merging proposed entities/edges that refer to the same real-world thing. Uses similarity over embeddings + property matching.

Relationship mapping — converting LLM-proposed relationships to canonical edges with stable IDs.

Commit — the final transactional stage that writes deduplicated nodes and edges to the graph and updates the source's status to committed.

Status flow — the lifecycle of a source: pending → indexing → indexed → extracting → extracted → committing → committed, plus error and mcp_extracting.

Citation — a back-reference from a generated answer to the chunk(s) that grounded it.

MCP — Model Context Protocol. Chaos Cypher exposes 30 MCP tools for AI clients (19 read, 11 write).

Operations queue / LLM queue — the two named worker queues. Operations (8 concurrent) handles indexing/dedup/commit; LLM (1 concurrent) handles extraction LLM calls.

Lexicon Hub — Chaos Cypher's hosted domain-template registry. (Preview — see banner on the Lexicon Hub pages.)

CCX — Chaos Cypher eXchange, the canonical backup/export format.

RAG — Retrieval-Augmented Generation. Chaos Cypher's chat surface uses RAG with hybrid (BM25 + vector) retrieval.

FTS5 — SQLite Full-Text Search version 5. The backend for keyword search.

BM25 — the ranking algorithm used by FTS5. Default field weights: label 3.0, properties 1.0, searchable_text 0.5.

RRF — Reciprocal Rank Fusion. The algorithm that merges BM25 and vector ranks into a single hybrid score.

MRL — Matryoshka Representation Learning. Used by the default embedding model (Qwen/Qwen3-Embedding-0.6B) to support truncated embeddings.