Glossary
Terms as this codebase uses them. When a word has a general industry meaning and a specific meaning here, the entry gives the specific one.
Platform & product
VeyraOps AI — the product/brand name of this repository and documentation.
AgentOps AI — the original internal name, still visible in the FastAPI title, frontend UI, agentops.db, and the AGENTOPS_* env vars. Same product (naming note).
Agents & tools
Agent — a stored configuration (name, type, system prompt, language, tone, enabled_tools) that owns all its data. A row in agents, not a running process (Agents).
Workflow agent — one of the three LLM roles inside the email pipeline (Analysis / Reply / Reviewer). A pipeline stage, not an agents row (terminology note).
Tool — a capability grant on an agent, stored as a comma-separated ID in enabled_tools and enforced by agent_has_tool (Tools). Canonical IDs: email_analyzer, task_creator, document_search, reply_generator.
Gate tool / enhancement tool — the two enforcement styles: a missing gate tool rejects the request (403); a missing enhancement tool degrades behavior silently (semantics).
System prompt — the agent-author-written instruction text, wrapped at runtime in the platform's identity/rules template (Prompts).
Execution & telemetry
AgentRun — the record of one chat exchange: input, output, latency, status (Observability).
ToolCall — the observability atom: one row per automated step, with success flag, latency, and (for costed steps) model/tokens/cost.
WorkflowRun / WorkflowStep — one pipeline execution and its per-stage trace records (Workflows).
Observability contract — the platform rule that every automated step logs a ToolCall and pipelines trace themselves (the contract).
Quality score — the Reviewer Agent's 0–1 verdict on a workflow's output, stored (as a string) on the run and averaged on the dashboard (Evaluation).
Estimated tokens / estimated cost — chars÷4 heuristics priced at gpt-4o-mini rates; monitoring signals, not billing data (Cost Monitoring).
RAG & documents
RAG (Retrieval-Augmented Generation) — answering from retrieved document context instead of model memory (RAG Overview).
Chunk — an ~800-character slice of an uploaded document (100-char overlap), stored in SQLite and embedded into Chroma (Ingestion).
Embedding — the 1536-dim vector representation of a chunk or query, from text-embedding-3-small.
Chroma — the local persistent vector store; one agent_documents collection for all agents (Database Overview).
vector_id — the deterministic Chroma record ID (agent_{a}_document_{d}_chunk_{i}) bridging SQLite and Chroma.
Score (retrieval) — Chroma's raw distance; lower is better (Retrieval).
Corrective RAG — the LangGraph pipeline that evaluates retrieval sufficiency before answering, refusing otherwise (Corrective RAG).
Retrieval evaluator — the temperature-0 LLM judge (has_answer/confidence/reason) gating generation (Evaluation).
Refusal string — the fixed sentence "I could not find this information in the uploaded documents.", produced by both the refusal node and the answer prompt's rules.
Workflow pipeline
Analysis Agent — stage 1; extracts summary/intent/priority/deadline/action items/suggested reply as strict JSON.
Task Agent — stage 2; deterministic Python that persists a Task per action item (no LLM).
Reply Agent — stage 3; drafts the outbound reply (requires reply_generator, else fallback).
Reviewer Agent — stage 4; the advisory quality judge.
normalize_due_date — the deadline sanitizer: keeps simple relative words, drops past ISO dates, passes other text through (helper docs).
Architecture & infrastructure
Router / Service / Model / Schema — the four backend layers: HTTP+orchestration, AI logic, persistence, validation (Backend).
BFF proxy — the Next.js route handler (app/api/agents/route.ts) that forwards browser writes to FastAPI (API Proxy).
Server component — a React component rendered on the server per request; the frontend's default (Data Fetching).
Startup migration — the boot-time ALTER TABLE mechanism in run_startup_migrations() (Migrations).
LangGraph / StateGraph — the graph-orchestration library and its state-machine class; used only for corrective RAG (Design Decisions).
uv — the Python package manager used for both the backend and the docs build.
Design tokens — the semantic color/typography/spacing classes the UI must use (Design System).