Blog Writing Agent (Researcher + AI Agent)

This repository contains a production-ready research-driven blog writing agent that turns a user topic into a publishable markdown blog with supporting images.

Overview

The system implements a conservative, token-aware pipeline that reliably produces a blog even when LLM rate limits are reached. High-level flow:

• Start: user supplies a topic and an as_of date.
• Router: quickly decides whether external research is needed and picks a mode (closed_book, hybrid, open_book).
• Research (optional): runs web searches (Tavily) and extracts a small, high-signal evidence pack.
• Orchestrator: creates a 4–6 task plan (title, audience, tone, tasks).
• Worker(s): sequentially generate section markdown for each task (serialized to avoid token bursts).
• Reducer (merge_content -> decide_images -> generate_and_place_images): merge sections, plan images, generate images, and write final markdown.
• End: final markdown file is written (in working directory) and images are saved to images/.

Key features

• Conservative token budgets and caps to avoid per-minute and per-day quota spikes.
• Groq LLM integration via langchain-groq with graceful fallback when daily TPD limits are hit (the app will still produce a blog + fallback visuals).
• Research via Tavily (optional; requires TAVILY_API_KEY).
• Free image generation using Pollinations (no API key required) with a local SVG fallback when image generation fails.
• Streamlit frontend for running the app interactively (bwa_frontend.py).
• Generates a .md file per blog and writes images to images/.

Tech stack

• Python 3.10+ (Conda environment: bwa)
• LangGraph (graph orchestration)
• LangChain + langchain-core
• Groq LLM via langchain-groq (primary LLM)
• Tavily search via langchain-tavily (web search)
• Pollinations image generation (https://image.pollinations.ai) for cost-free images
• Streamlit frontend (bwa_frontend.py)
• Utilities: pydantic, python-dotenv, pandas

Files of interest

• bwa_backend.py — main graph and node implementations (router, research, orchestrator, worker, reducer).
• bwa_frontend.py — Streamlit UI for running the agent and previewing output.
• requirements.txt — recommended Python packages (see langchain-tavily addition).
• images/ — generated images are written here at runtime.
• screenshots/ — example visuals used in this README (embedded below).

Environment & usage

1. Create or activate the bwa conda environment.

conda activate bwa
pip install -r requirements.txt

2. Add a .env file in the project root with the following (if you use Groq/Tavily):

GROQ_API_KEY=your_groq_api_key_here
TAVILY_API_KEY=your_tavily_api_key_here  # optional, for open_book/hybrid modes

3. Run the Streamlit frontend:

streamlit run bwa_frontend.py

Outputs

• A markdown file is written for each generated blog (filename derived from the title) in the repository root.
• Images are written into the images/ folder. If Pollinations returns non-image content or is unreachable, the agent writes an auto-generated SVG fallback visual for each planned image so the markdown remains complete.

Design notes

• The router is intentionally lightweight and low-token to decide whether research is necessary.
• Workers are executed sequentially (worker_loop_node) to avoid parallel LLM calls and reduce tokens-per-minute usage.
• When the Groq daily token hard cap (TPD) is reached, the system falls back to an internal plan + section renderers so you still get a coherent, actionable blog and images.

Screenshots

Below are the screenshots included in the screenshots/ folder (displayed in serial order):

1. Image 1

2. Image 2

3. Image 3

4. Image 4

5. Image 5

6. Image 6

7. Image 7

Troubleshooting

• If you see repeated Conda warnings about C:\\Users\\<you>\\.condarc, fix or remove that file — it does not affect the app but produces noise in terminal output.
• If Groq returns 429/day-limit errors, the agent will fall back and continue producing content; to avoid this, upgrade your Groq plan or reduce usage.

Contributing

If you want to improve planning prompts, image prompts, or add alternative image providers, modify bwa_backend.py and adjust the prompt templates and the _pollinations_generate_image_bytes implementation.

License & contact

This project is provided as-is. For questions or improvements, open an issue or contact the maintainer.