Glossary — Lawyer Assistant

Use this as a companion to AGENT_KNOWLEDGE_BASE.md (Q&A) and README.md (index).

A–C

• AsyncSqliteSaver — LangGraph checkpointer that persists agent thread state to data/conversations.db on disk (falls back to MemorySaver). Note the checkpointer is project-independent — it does not move with the workspace. → BACKEND.md §4
• BGE-M3 — the multilingual embedding model (BAAI, 1024-dim, L2-normalized) used to turn text chunks and queries into vectors for dense search. → BACKEND.md §5
• BGE-Reranker-v2-M3 — the cross-encoder model that re-scores retrieved candidates to pick the best top-5. → BACKEND.md §5
• BM25 — the classic keyword (sparse) ranking algorithm used for the sparse retrieval leg, backed by a prebuilt inverted index in <project>/workspace/bm25_index/ when a workspace is active, else data/bm25_index/. → BACKEND.md §5, DATA_FLOW.md
• Bypass edge — a temporary particle edge created when a pipeline node is removed, rewiring its incoming sources straight to its outgoing targets so the flow stays connected; tagged bypass-<nodeId>- so re-enabling finds it. → PIPELINE_EDITOR.md §3
• Chain-of-thought (thinking) — the LLM's intermediate reasoning tokens streamed to the UI (before tools are called) and shown as a collapsible "thinking" block. → FRONTEND.md §6
• Checkpointer — the LangGraph component that stores conversation graph state (threads), enabling human-in-the-loop pause/resume across HTTP requests. → BACKEND.md §4b
• ChromaDB — the local vector database (collection legal_chunks, cosine distance, HNSW index) storing chunk embeddings and metadata. With a workspace active it lives in <project>/workspace/chroma_db, else data/chroma_db. → BACKEND.md §5
• Chunk — a 480-token slice of a document (sections never split) that gets embedded and stored; the atomic unit of retrieval. → DATA_FLOW.md
• Citation — the source reference (file, page, section) attached to a retrieved chunk and shown inline in answers; clicking opens the SourceViewerModal. → FRONTEND.md §2, §5
• Conversation ID — the thread identifier passed with every chat request (client-supplied, falling back to a server-generated UUID) so agent state and history follow the conversation. → BACKEND.md §3
• CORS — backend middleware allowing the Vite dev origin (http://localhost:5173) and file:// to call the API. → BACKEND.md §1

D–H

• Dense retrieval — semantic search over ChromaDB embeddings (vector similarity), the dense pipeline node. → BACKEND.md §5, PIPELINE_EDITOR.md §2
• Docling — the document-parsing library (with OCR) used during ingestion to extract text, tables, and structure from PDFs/images. → BACKEND.md §5, DATA_FLOW.md
• Electron — the desktop shell that hosts the app: launches the Python backend and Vite dev server, owns the window, exposes IPC. → ARCHITECTURE.md §1, §3
• Embedding — a 1024-dim numeric vector representing text semantics, produced by BGE-M3; cosine similarity between vectors drives dense search. → BACKEND.md §5
• FastAPI — the Python web framework serving the app's HTTP + SSE API on port 8765 (backend/main.py). → BACKEND.md §1
• Flag — a compliance finding from the playbook scan (severity, rule id, page, reason, clause text); statuses open → resolved/dismissed/ escalated, persisted to SQLite. → BACKEND.md §6
• HITL (human-in-the-loop) — the approval step where the LangGraph agent pauses via interrupt() before executing a tool and waits for the user to approve/reject; resumed through /api/chat/resume. → BACKEND.md §4b
• HNSW — the approximate nearest-neighbor graph index inside ChromaDB that makes dense search fast. → BACKEND.md §5
• Hybrid search — combining dense (vector) and sparse (BM25) results; the default mode merges them (RRF / boost-only / combine-and-dedup strategies). → BACKEND.md §5

I–P

• Mode — the chat request mode: rag (retrieve + answer, default), retrieval_only (search only, no LLM), or direct (LLM only, no retrieval). → BACKEND.md §2
• Ingestion — the pipeline that turns raw files into indexed chunks: parse (Docling/OCR) → chunk → embed → store (ChromaDB + BM25). → DATA_FLOW.md, BACKEND.md §5
• Intent classifier / router — the default RAG orchestrator (router_intent.py): classifies the message intent, writes a short plan, detects the tool from the plan, executes it, and writes the answer. → BACKEND.md §4a
• IPC — Electron's inter-process communication used for window controls and backend health/port queries. → ARCHITECTURE.md §1
• LangGraph — the orchestration framework powering the agentic router's StateGraph (agent → approval → tools) and checkpointer. → BACKEND.md §4b
• MemorySaver — the in-memory fallback checkpointer used when the SQLite checkpointer can't be created. → BACKEND.md §4b
• Ollama — the local LLM runtime (http://localhost:11434) providing generation for RAG answers, relevance checks, and scanning. → BACKEND.md §7
• Particle edge — the animated ReactFlow edge type used throughout the pipeline editor to visualize data flowing between nodes. → PIPELINE_EDITOR.md §1
• Playbook — a JSON rule set (backend/legal_retrieval/playbook/playbooks/, e.g. default.json) defining clause types and compliance rules the scanner checks documents against. → BACKEND.md §6
• Playbook scan — the clause-level compliance check: chunk → classify clause type → check rules → emit flags → persist. → BACKEND.md §6

R–S

• RAG (retrieval-augmented generation) — answering a question by first retrieving relevant document chunks, then having the LLM write an answer grounded in those chunks with citations. → BACKEND.md §3, §5
• ReAct — the "Reasoning + Acting" agent loop (legacy router.py) where the LLM decides which tool to call at each step. → BACKEND.md §4b
• Reranker — the cross-encoder (BGE-Reranker-v2-M3) that re-scores the ~50 retrieved candidates and keeps the top 5; can be skipped (skip_rerank). → BACKEND.md §5
• RRF (Reciprocal Rank Fusion) — the rank-based merge strategy used in hybrid search to combine dense and sparse rankings. → BACKEND.md §5
• Sparse retrieval — BM25 keyword search over the inverted index; the sparse pipeline node. → BACKEND.md §5
• SSE (Server-Sent Events) — the text/event-stream protocol the backend uses to stream chat tokens, statuses, tool calls, node progress, and scan progress to the UI in real time. → BACKEND.md §3, FRONTEND.md §2
• Streaming event — one typed SSE payload (e.g. token, status, tool_call, node_start, done); the frontend switches on these types. → BACKEND.md §3

T–Z

• Thread / thread_id — the LangGraph conversation thread key (same as the conversation id) used to pause, store, and resume agent state. → BACKEND.md §4b
• Tool — a function the agent can call: search_documents, verify_relevance, scan_document, ingest_file. → BACKEND.md §4c
• Top-k — how many results are requested/returned: 50 candidates retrieved, 5 returned after reranking (top_k_final). → BACKEND.md §5
• Upload — saving an attached file into the active project folder via /api/upload so it can be ingested/scanned with the next message (a workspace must be selected first). → BACKEND.md §2
• Vector store — ChromaDB's persistent storage of chunk embeddings + metadata (file, page, section) used for dense retrieval. → BACKEND.md §5
• Warmup — the background thread at backend startup that preloads the embedder, reranker, LLM, and classifier references so the first query is fast. → BACKEND.md §1
• Working set — the pinned/selected sources panel that collects documents of interest during a session. → FRONTEND.md §5
• Zustand — the lightweight React state store persisting conversations (freebuff-chat-storage) and holding all streaming/approval state. → FRONTEND.md §3