Verified Dev Copilot (LangChain + optional LangGraph)

Overview

A developer-focused chatbot agent that answers technical questions using:

• DuckDuckGo web search
• StackOverflow search (StackExchange API)
• Official documentation discovery (site-biased search)
• Page fetching for grounded excerpts
• Human approval loop before finalizing

Designed as a portfolio-ready “agentic research” demo for Upwork profiles.

Flow Diagram

flowchart TD
    A[CLI Input] --> B{Engine}
    B -->|basic| C[Analyze Question]
    B -->|langgraph| D[LangGraph StateGraph]
    D --> C
    C --> E{Need Clarification?}
    E -->|yes| F[Collect Human Info]
    E -->|no| G{Use Tools?}
    F --> G
    G -->|yes| H[Search + Fetch Sources]
    G -->|no| I[Draft Answer]
    H --> I
    I --> J[Return JSON: answer + sources]

    classDef entry fill:#E8F0FF,stroke:#2B59C3,stroke-width:1px,color:#0B1A33;
    classDef decision fill:#FFF4CC,stroke:#D08C00,stroke-width:1px,color:#3A2B00;
    classDef process fill:#EAF7EE,stroke:#2E7D32,stroke-width:1px,color:#0B2B12;
    classDef output fill:#F3E8FF,stroke:#6A1B9A,stroke-width:1px,color:#2A0A3D;

    class A entry;
    class B,E,G decision;
    class C,D,F,H,I process;
    class J output;

Loading

What it Returns

JSON with:

• answer: final response (includes a Sources: section)
• sources: URLs found in the answer (filtered to fetched URLs)

Demo

python app/agent.py --question "Why does pip say SSL: CERTIFICATE_VERIFY_FAILED on macOS?"

Example output:

{
  "answer": "Try these steps on macOS:\n1. Install/refresh certs ...\n2. Upgrade pip and certifi ...\n3. Check corporate proxy or custom CA ...\n\nSources:\n- https://pip.pypa.io/en/stable/user_guide/#ssl\n- https://docs.python.org/3/library/ssl.html",
  "sources": [
    "https://pip.pypa.io/en/stable/user_guide/#ssl",
    "https://docs.python.org/3/library/ssl.html"
  ]
}

Reliability & Guardrails

• Human approval gate before any web/tools research (unless --no-approval is set).
• Citations must come from fetched URLs; sources list is filtered to those URLs.
• Clarifying questions are asked only when needed to avoid wrong advice.

Configuration

Variable	Default	Purpose
LM_STUDIO_BASE_URL	http://localhost:1234/v1	Local OpenAI-compatible endpoint
LM_STUDIO_API_KEY	lm-studio	API key for the local endpoint
LM_STUDIO_MODEL	qwen/qwen-3-1.7b	Model name for the local endpoint

CLI flags:

• --engine langgraph|basic
• --max-sources, --max-page-chars, --max-context-chars
• --no-approval

Project Layout

• app/agent.py: CLI entrypoint (basic engine + optional LangGraph engine).
• app/tools.py: Search + fetch tools (DuckDuckGo, StackOverflow, docs search).

Run

python app/agent.py --question "How do I fix pip SSL errors on macOS?"

Optional:

python app/agent.py --engine basic --question "How do I fix pip SSL errors on macOS?"

Interactive CLI

python app/agent.py --interactive

Notes

• DuckDuckGo + StackOverflow require outbound network access.
• The default LLM config targets a local OpenAI-compatible endpoint: LM_STUDIO_BASE_URL="http://localhost:1234/v1" (e.g., LM Studio).
• Set LM_STUDIO_API_KEY if your local endpoint requires a key (default: lm-studio).
• Override the model with LM_STUDIO_MODEL (default: qwen/qwen-3-1.7b).
• .env is supported if you install python-dotenv.
• The agent asks for human input when it needs missing details (OS, versions, errors).
• The agent asks for approval only when it wants to use web/tools for research.
• Install questionary for nicer interactive prompts (arrow-key selection + input highlight).
• Install rich for panelled markdown output in interactive mode.

Limitations

• Web search and StackOverflow APIs depend on external network access.
• HTML parsing can be brittle if providers change their markup.
• Output quality depends on the local model and prompt adherence.