Glossary

Project terminology, alphabetical. Each term is defined as it is actually used in the codebase; see the linked concept pages for depth.

Anti-fabrication guard

The orchestrator step that rewrites an answer to "not found" when the structured channel returns no found fact. Enforced in control flow (agent/agent.py), not in a prompt. See Anti-fabrication.

Chunk store

The sqlite-backed store for narrative document chunks (retrieval/chunking/chunk_store.py). Each chunk carries doc_id + source_locator (NOT NULL) plus metadata (title, entity, period, sensitivity, …). Current rows also persist parent_id, heading, window_text, and parent_locator for generation-only small-to-big/window expansion.

Economy retrieval mode

The economy / bm25 / lexical service preset. It constructs neither an embedding backend nor a vector store and performs lexical retrieval only.

Clarification gateway

The step (clarify_scope, agent/intent.py) that decides — by a deliberate asymmetry — whether to ask first (missing metric), answer with surfaced assumptions (missing entity/period), or refuse (out-of-scope entity).

Composite (route)

A question that is both numeric and narrative (a recognized metric + a narrative cue). The agent runs the structured path, then appends a narrative attribution section.

Dual channel

The structured numeric path + the narrative RAG path, run deterministically and — for composite questions — compared and merged. See Dual channel.

dim_key

The canonical sorted-JSON natural key over a fact's identity dimensions (metric, entity, channel, period). The UNIQUE upsert conflict key in the fact store; storage-only, recomputed on write — never a Fact field.

fact_metric

The sqlite table holding every structured numeric fact, one row per metric × entity × period × channel, with full lineage (source_doc_id, source_locator). Backed by storage/fact_store.py.

FAQ short-circuit

An SME-vetted question→answer cache (service/faq/) that bypasses the LLM, sitting in front of the anti-fabrication guard, behind conservative exclusions. See FAQ short-circuit.

found / not_found / unrecognized_param

The tri-state returned by the query_metric tool: a value exists (found), no matching row (not_found), or a parameter could not be normalized to a controlled code (unrecognized_param). The structured channel never returns a guess.

Glossary (synonym normalization)

The mapping layer (common/glossary.py) that normalizes ZH/EN/abbreviation synonyms to controlled metric / entity / period codes. Returns None on anything unrecognized rather than guessing — the entry point to the structured channel.

Home company / CompanyProfile

The deployment's own company identity, metrics, entities, and competitor list, loaded from config (config/company.example.toml). No company is hardcoded anywhere; "ACME" is only the fictional demo profile.

Intent slots

The four parsed slots — metric, entity, period, channel — extracted from a question by the intent parser and used for routing and querying (agent/intent.py).

History turn

A (role, text) tuple supplied to answer_question(history=...) or /v1/ask. History is LLM generation context only; it is excluded from intent/security decisions, retrieval queries, and evidence.

IR / StyledGrid

The frozen intermediate representation that documents are extracted into (style- and color-aware) before ingestion into the fact and chunk stores.

IntentParser (Protocol)

The pluggable seam for intent extraction (ADR 0010). Default is the zero-LLM, config-driven RuleIntentParser. Security is not pluggable — the security gate re-derives scope from the raw question independently of the parser.

Listwise rerank

An optional second-pass reranker (retrieval/rerank/listwise_rerank.py) that sends the top candidates to an LLM judge for relevance ordering, falling back to RRF order on any failure. RESTRICTED candidates never enter the judge prompt.

MockProvider

The deterministic, offline LLMProvider implementation (agent/llm_provider.py). Lets the core run fully offline with no API key — the default for demos and tests.

Narrative channel

The RAG path for "why / what happened" questions: hybrid retrieve → optional rerank → synthesis with citations. Default chain is pure BM25 + RRF (no vector backend wired).

Metadata filter

A validated narrowing predicate over chunk metadata. Supported operators are eq, ne, in, nin, gt, gte, lt, lte, and between; filtering preserves incoming result order.

Multi-index retrieval

Per-library retrieval followed by RRF fusion. Each result retains library_id provenance; routing failure falls back to all configured libraries.

Parent-child expansion

Small-to-big retrieval where the child determines rank and citation, while persisted parent/window text becomes separate generation-only prompt_text. Restricted children and their windows are dropped together.

Provenance

The source_doc_id (or doc_id) + source_locator carried by every fact, chunk, and answer. Carried end-to-end, never dropped. See Provenance.

Protocol seam

A typed Protocol boundary at which an external dependency is injected (LLM provider, embedding backend, listwise judge, OCR backend, narrative retriever, task queue). The core imports zero SDKs and depends only on these abstractions.

Preview protocol v1

A display-only workflow projection containing version, redacted nodes/edges, and deterministic geometry. It contains no prompts, variables, providers, credentials, or executable configuration.

query_metric

The function-calling tool of the structured channel (agent/query_tools.py). Normalizes its parameters via the glossary, queries fact_metric, and returns the tri-state result with lineage — never a fabricated number.

Reciprocal Rank Fusion (RRF)

The rank-fusion method combining BM25 and vector rankings in hybrid retrieval (rrf_fuse, k=60); each ranking contributes 1 / (k + rank) to an item's fused score.

RESTRICTED isolation

The invariant that RESTRICTED-tier content is stripped at two exits (retrieval/link and retrieval/rerank) before it can reach a prompt. See RESTRICTED isolation.

Review queue

The human (SME) review-queue state machine for low-confidence or conflicting ingested items (ingestion/review/). Distinct from the async job queue; a fact's review_status (e.g. pending, approved, rejected, blocked) gates query visibility.

Security gate

The deterministic, never-pluggable front door (agent/security_gate.py, ADR 0010) that detects external/competitor entities by longest-match, masks them with equal-length spaces, and refuses out-of-scope questions before any tool / retriever / LLM call.

Sensitivity tier

The per-chunk classification (common/sensitivity.py): INTERNAL by default, escalated to RESTRICTED by deterministic, config-driven, fail-safe signals.

Structured channel

The deterministic numeric path: glossary normalization → query_metric over the fact_metric store → answer rendered from the fact value, with lineage. Never invents a number.

Workflow catalog

Package-owned templates used by the scaffold command and website. It contains 7 curated workflows and 993 generated descriptors across 39 industries, 27 use cases, and five archetypes. Upstream Dify/n8n pages are source metadata only; their workflow configuration is not redistributed or downloaded at runtime.

Workflow scaffold

A pure configuration operation that explicitly selects or matches a catalog template, otherwise generating a fixed Dify 0.6 start → llm → end graph. It returns configuration but neither writes nor executes it; the CLI separately performs safe file output.