Add README.md

This commit is contained in:
2026-06-04 16:16:27 +00:00
commit f8457690fc
+67
View File
@@ -0,0 +1,67 @@
# LangGraph DraftCritiqueRewrite Agent
## Overview
This project implements a small LangGraph agent that:
1. Generates a short answer (510 sentences) to a user question.
2. Critiques the draft for completeness, concreteness and absence of filler.
3. If the critique says *needs_revision*, rewrites the draft incorporating the critique.
4. Loops until the verdict is *ok* or a maximum of two rewrites is reached.
5. Provides a simple CLI entry point.
The implementation is fully selfcontained and requires only a modern Python interpreter (3.10+).
## Installation
```bash
# Create and activate a virtual environment (optional but recommended)
python -m venv .venv
source .venv/bin/activate # On Windows: .venv\Scripts\activate
# Install dependencies
pip install -r requirements.txt
```
**Dependencies**
- `langgraph`
- `langchain-openai`
- `langchain-ollama`
## Configuration
The agent uses an LLM via the `langchain_openai.ChatOpenAI` wrapper. It supports both OpenAI and Ollama providers.
| Variable | Description | Example |
|----------|-------------|---------|
| `OPENAI_API_KEY` | API key for OpenAI. | `export OPENAI_API_KEY=sk-…` |
| `OLLAMA_BASE_URL` | Base URL for a local Ollama instance. | `export OLLAMA_BASE_URL=http://localhost:11434/v1` |
If neither variable is set the agent falls back to the default model `gpt-4o-mini` (this will require network access to OpenAI).
## Usage
Run the CLI entry point:
```bash
python agent.py
```
You will be prompted to type a question. The agent will then display the final draft, the critique, and the verdict.
### Example
```
$ python agent.py
Enter your question: What is the capital of France?
Final Result:
Draft: The capital of France is Paris.
Critique: ok
Verdict: ok
```
## Development Notes
- The graph uses `END` from `langgraph` for terminal states.
- The `reflect` node parses a JSON response, falling back to a default verdict if parsing fails.
- The `rewrite` node increments the round counter.
- All LLM calls are wrapped in `try/except` blocks to surface errors clearly.
## License
MIT License.