ArcLight Copilot

Illuminating research with agentic AI.
ArcLight Copilot is a demo-ready, portfolio-grade agentic research assistant built with LangChain and Azure OpenAI, featuring planning, tool use (web/RAG), reflection, and a clean Streamlit UI.

Why this repo? Most demos are chat wrappers. ArcLight plans tasks, uses tools, cites sources, and self-checks answers—showcasing real agent patterns and Azure integration.

---

✨ Highlights

• Agentic flow: Planner → Researcher (RAG/tool use) → Reviewer (reflection) → Final Answer
• Azure-first: Azure OpenAI (LLM & embeddings), optional Azure AI Search for RAG
• Streamlit UI: one-file run, live step log, exportable run trace
• Quality: tests, CI (ruff/black/pytest), Dockerfile, DevContainer
• Safe defaults: Runs in a demo mode without API keys; full features light up when env vars are set

---

🧱 Architecture (at a glance)

User
 └─ Streamlit UI
     └─ ArcLight Orchestrator
         ├─ Planner Agent (LangChain Tools Agent)
         ├─ Researcher (RAG chain + tools)
         ├─ Reviewer (critique & retry)
         └─ Finalizer (answer + citations + action log)
Tools:
  - Web Search (stub or Tavily/SerpAPI)
  - Azure AI Search (optional)
  - GitHub Issues (stub)
  - Python Sandbox (restricted)
Memory:
  - Local JSON store (dev) or plug in Cosmos DB

---

🚀 Quickstart

1) Create a virtual env & install

python -m venv .venv && source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

2) Configure environment

cp .env.example .env
# Fill Azure settings if you want live LLM and RAG

3. Install application:

pip install -e .

4) Run

streamlit run app/ui_streamlit.py

No keys? No problem. ArcLight will run in Demo Mode with canned agent outputs so you can see the UX and flow.

---

🔧 Environment Variables

Copy .env.example and fill as needed.

Required for live LLM

• AZURE_OPENAI_ENDPOINT
• AZURE_OPENAI_API_KEY
• AZURE_OPENAI_DEPLOYMENT (e.g., gpt-4o-mini)
• AZURE_OPENAI_API_VERSION (e.g., 2024-06-01)

Optional for RAG

• AZURE_SEARCH_ENDPOINT
• AZURE_SEARCH_KEY
• AZURE_SEARCH_INDEX

---

📂 Repo Structure

ArcLight-Copilot/
├─ app/
│  └─ ui_streamlit.py
├─ arclight/
│  ├─ agents/
│  │  ├─ planner.py
│  │  ├─ researcher.py
│  │  └─ reviewer.py
│  ├─ chains/
│  │  └─ rag_chain.py
│  ├─ tools/
│  │  ├─ web_search.py
│  │  ├─ github_issues.py
│  │  └─ py_sandbox.py
│  ├─ models/
│  │  └─ llm.py
│  ├─ memory/
│  │  └─ conversation_store.py
│  └─ config.py
├─ tests/
│  ├─ test_tools.py
│  └─ eval_tasks/
│     └─ tasks.yaml
├─ .github/workflows/ci.yml
├─ .devcontainer/devcontainer.json
├─ Dockerfile
├─ requirements.txt
├─ .env.example
├─ .gitignore
└─ README.md

---

🖥️ Streamlit UI

• Left sidebar: Model/feature status, demo-mode toggle, export trace
• Main: Prompt box → live “Agent Log” → final answer with citations
• RAG panel: Shows retrieved docs (when Azure AI Search is configured)

---

🧪 Tests & Eval

• tests/test_tools.py – sanity checks for tools (no external calls in CI)
• tests/eval_tasks/tasks.yaml – sample eval tasks for regression checks

Run tests:

pytest -q

---

🔐 Security & Safety Notes

• Python sandbox is restricted (builtins off, tiny allowlist).
• Web search tool is a stub by default—swap with Tavily/SerpAPI for production.
• Do not store keys in code. Use .env or your secret manager.

---

🗺️ Roadmap

• Ingestion script to push PDFs/notes into Azure AI Search
• Richer reviewer with structured hallucination scoring
• Next.js UI option
• Trace export as shareable JSON (replay in UI)

---

🙌 Credits

Built using LangChain, Streamlit, and Azure services. MIT Licensed.