langchain-agentic-patterns

# langchain-agentic-patterns > **Faber** — a production AI agent built from LangGraph v1 and LangChain v1 primitives. Deep research, voice, RAG, and MCP tools, composed on a custom middleware harness. <!-- Hero gif / screenshot goes here once Ship 1 is live --> > **Status:** 🚧 Phase 1 — Foundation Chatbot (building toward Ship 1). See [ROADMAP.md](ROADMAP.md). ## What it does - **Text or voice research queries** — ask a question by chat or by speaking - **Plan generation with clarification** — the agent scopes the question, asking for clarification when needed (human-in-the-loop) - **Parallel multi-agent research** — an orchestrator dispatches isolated sub-agents for web search and document retrieval - **Synthesis with citations** — a writer agent composes a grounded report with inline source citations - **Persistent memory** — user preferences and prior topics carry across sessions - **Extensible via MCP** — any Model Context Protocol server becomes a connector ## Quick start > 🚧 **Not runnable yet.** There is no application code or container setup in the repo yet. > The one-command start below is the *target* for **Ship 1** (Week 5), when > `docker-compose.yml`, the API, and the UI land. ```bash # Planned for Ship 1 — does not work yet: cp .env.example .env # fill in your API keys docker compose up ``` ## Architecture A three-layer stack — LangGraph v1 (runtime), LangChain v1 (framework), and a custom agent harness assembled from primitives. The harness composes a middleware stack (planning, memory, filesystem, sub-agents, summarization, HITL, tool-call patching, guardrails) over a set of specialized agents. See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the full overview and module map. ## Tutorial notebooks Eleven notebooks are **planned** — one per module, each explaining *why* a pattern exists before *how* to build it. None are published yet; see [ROADMAP.md](ROADMAP.md) for the schedule. | # | Notebook | Module(s) | |---|----------|-----------| | 01 | `01_chat_agent_foundation.ipynb` | M1 — Chat Agent Foundation | | 02 | `02_middleware_foundations.ipynb` | M2 — Middleware Foundations | | 03 | `03_rag_pipeline.ipynb` | M3 — RAG Pipeline | | 04 | `04_memory.ipynb` | M4 — Memory | | 05 | `05_hitl.ipynb` | M5 — Human-in-the-Loop | | 06 | `06_sub_agents.ipynb` | M6 — Sub-Agents | | 07 | `07_deep_research.ipynb` | M7 — Multi-Agent Deep Research | | 08 | `08_mcp_integration.ipynb` | M8 — MCP Integration | | 09 | `09_chat_app.ipynb` | M9 — Chat App (API + UI + sessions) | | 10 | `10_eval_and_deploy.ipynb` | M10 — Eval + Deploy | | 11 | `11_voice_interface.ipynb` | M11 — Voice Interface | ## Benchmarks | Benchmark | Score | Model | Date | |-----------|-------|-------|------| | Deep Research Bench | _pending (Ship 2)_ | — | — | | Internal eval (30q) | _pending (Ship 2)_ | — | — | ## Deployment Planned for Ship 1 / Ship 3: a Docker Compose setup and a hosted demo, to be documented in `docs/DEPLOYMENT.md` (not written yet). ## Tech stack (planned) Target stack for the full build. Only **LangGraph + LangChain** are installed so far; the rest lands per [ROADMAP.md](ROADMAP.md). | Layer | Component | |-------|-----------| | Runtime | LangGraph v1 (1.x) | | Framework | LangChain v1 (1.x) | | Language | Python 3.12 | | Package manager | uv | | Vector DB | ChromaDB (dev) → pgvector (prod) | | Web search | Tavily | | API | FastAPI | | Frontend | Web app (framework TBD — decided later via ADR) | | Voice | Deepgram (STT) + ElevenLabs (TTS) | | Tracing | LangSmith | ## Acknowledgments This project draws on three references in agent engineering: - **Concept reference — [*Agentic Design Patterns* (Antonio Gulli)](https://link.springer.com/book/10.1007/978-3-032-01402-3)**: the agentic patterns and the reasoning behind them. - **Architectural reference — [LangChain Deep Agents](https://github.com/langchain-ai/deepagents)**: middleware stack, BackendProtocol abstraction, sub-agent isolation, auto-summarization patterns. - **Product reference — [Onyx](https://www.onyx.app/)**: deep research workflow design, citation UX, feature scope. Studied via blog posts and product docs; no code copied. Implementation in this repository is the author's own, built directly on LangGraph and LangChain v1 primitives. Patterns are derived from internalized study, not from copied code. ## License Apache 2.0 — see [LICENSE](LICENSE).