BlogAgent

A LangGraph + Streamlit blog generation app that plans, researches, writes, and exports technical blog posts.

What This Project Does

• Builds a structured blog plan with section tasks and writing constraints.
• Optionally performs web research with Tavily and normalizes evidence.
• Writes each section with an LLM worker node.
• Merges sections into final markdown.
• Optionally decides and generates technical images.
• Saves output markdown to generated_blogs/ and images to generated_blogs/images/.
• Provides a Streamlit UI with multi-chat history and result tabs.

Project Structure

• blogAgent.py: LangGraph backend pipeline, Pydantic schemas, Tavily research, planning/writing, image generation, and app graph compile.
• frontend.py: Streamlit UI, multi-chat sessions, invoke flow, tabs for Plan/Evidence/Markdown/Images/Logs.
• requirement.txt: Python dependencies.
• generated_blogs/: Generated markdown outputs.
• generated_blogs/images/: Generated image assets.
• .env: Environment variables for API keys (not committed).

Architecture (LangGraph)

Main graph:

1. router
2. research (conditional)
3. orachestrator
4. worker (fanout for section tasks)
5. reducer subgraph

Reducer subgraph:

1. merge_content
2. decide_images
3. generate_andplace_image

Output state includes final markdown and intermediate artifacts like plan, evidence, and image specs.

Requirements

• Python 3.10+ recommended
• API keys:
  • MISTRAL_API_KEY (required for ChatMistralAI)
  • TAVILY_API_KEY (required for web research mode)
  • GOOGLE_API_KEY (optional, used only when image generation is requested)

Setup

1) Create and activate a virtual environment

Windows PowerShell:

python -m venv venv
(Set-ExecutionPolicy -Scope Process -ExecutionPolicy RemoteSigned) ; (& .\venv\Scripts\Activate.ps1)

2) Install dependencies

pip install -r requirement.txt

If your requirements file encoding causes issues, open and save requirement.txt as UTF-8, then re-run install.

3) Configure environment variables

Create a .env file in the project root:

MISTRAL_API_KEY=your_mistral_key
TAVILY_API_KEY=your_tavily_key
GOOGLE_API_KEY=your_google_key

Notes:

• GOOGLE_API_KEY is optional.
• If image generation fails, the app still produces markdown and inserts a failure note where image placeholders exist.

Run the App

Start Streamlit:

streamlit run frontend.py

Open the local URL shown in terminal.

Streamlit UI Overview

Sidebar:

• Chats: Create a new chat and switch between conversations.
• Generate:
  • Topic input
  • As-of date
  • Generate Blog button

Main area:

• Chat History
• Tabs:
  • Plan
  • Evidence
  • Markdown Preview
  • Images
  • Logs

Behavior:

• Backend runs once per Generate click.
• Generate button is disabled while a request is in progress.
• Each chat stores its own output/history/logs.

Outputs

• Markdown files: generated_blogs/<safe_title>.md
• Image files: generated_blogs/images/
• Markdown image links are written as images/ so they resolve from generated_blogs context.

Programmatic Backend Usage

You can call the compiled graph directly:

from blogAgent import app

out = app.invoke({
    "topic": "Your topic",
    "mode": "",
    "needs_research": False,
    "queries": [],
    "evidence": [],
    "plan": None,
    "as_of": "2026-04-19",
    "recency_days": 7,
    "sections": [],
    "merged_md": "",
    "md_with_placeholders": "",
    "image_specs": [],
    "final": "",
})

print(out.get("final", ""))

There is also a helper:

from blogAgent import run
print(run("Topic")["final"])

Troubleshooting

Evidence appears empty or malformed

• Ensure TAVILY_API_KEY is set.
• The backend includes a deterministic fallback when LLM research JSON parsing fails.

Source domain is blank

• Source is derived from URL domain.
• If a result has no valid URL, it is skipped.

Task flags look wrong

• Task schema requires:
  • requires_research
  • require_citations
  • require_code
• Planning prompt explicitly instructs the model to set these.

No images generated

• Check GOOGLE_API_KEY.
• Image generation errors are handled gracefully and do not block markdown output.

Notes

• The filename is requirement.txt (singular), not requirements.txt.
• The output folder generated_blogs/ is git-ignored in this repo.