city2facebook

Generate AI-powered summaries of Denton, TX city government meetings and publish them to a Facebook page.

How It Works

1. Scans a meeting archive containing metadata.json, whisper_transcript.txt, and agenda.pdf per meeting
2. Sends the full transcript + agenda to a local LLM (llama.cpp) for structured summarization
3. Formats the summary as a Facebook post with key decisions, community impact, and relevant hashtags
4. Publishes to a Facebook Page via the Graph API, tracking posted meetings to avoid duplicates

Setup

# Clone the repo
git clone https://github.com/yourusername/city2facebook.git
cd city2facebook

# Create and activate a virtual environment
python3 -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r city2facebook/requirements.txt

# Copy example config and fill in your values
cp city2facebook/config.example.json city2facebook/config.json
# Edit config.json with your archive path, LLM details, and Facebook credentials

Facebook Page Token

To get a valid page access token:

1. Go to Facebook Graph API Explorer
2. Select your app, click "Get Token" → "Get User Access Token"
3. Ensure pages_manage_posts and pages_read_engagement permissions are checked
4. Generate the user token, then call /me/accounts?fields=id,access_token in the explorer
5. Copy the access_token from your page's entry in the response — this is the page token
6. Also note the page id from the same response
7. Place both in config.json

The page token lasts ~60 days. When it expires, repeat the process.

LLM Server

The tool expects a local LLM running the OpenAI-compatible API (e.g., llama.cpp):

./server -m /path/to/model.gguf --port 8000

Update llm_base_url, llm_model, and llm_key in config.json accordingly.

Usage

List meetings in the archive

python city2facebook/main.py list
python city2facebook/main.py list --search "zoning"
python city2facebook/main.py list --body "City Council"
python city2facebook/main.py list --limit 20

Process a single meeting

Dry run (generate summary, don't post):

python city2facebook/main.py process --meeting-id 108037 --dry-run

Interactive mode (review before posting):

python city2facebook/main.py process --meeting-id 108037
# Then choose: [p]ost, [s]kip, or [e]xport draft to file

Auto-post (generate and publish immediately):

python city2facebook/main.py process --meeting-id 108037 --auto

Force reprocess (re-generate and re-post even if already in the posted log):

python city2facebook/main.py process --meeting-id 108037 --auto --force

Verbose logging

Add -v to any command for debug-level output:

python city2facebook/main.py -v process --meeting-id 108037 --dry-run

Custom config

Use --config to point to a different config file (e.g., for local dev):

python city2facebook/main.py --config config.dev.json process --meeting-id 108037 --dry-run

Archive Structure

Expected directory layout per meeting:

archive/
  bag-108037/
    data/
      metadata.json          # Meeting title, date, chapters
      whisper_transcript.txt # Full meeting transcript
      agenda.pdf             # Meeting agenda document
      video.mp4              # (optional, ignored by the tool)

Configuration

Key	Description
archive_dir	Path to the root archive directory containing bag-* folders
llm_base_url	Base URL of the OpenAI-compatible LLM server
llm_model	Name of the model to use
llm_key	API key for the LLM server
fb_page_id	Facebook Page numeric ID
fb_token	Facebook Page Access Token
posted_log	Path to the JSONL file tracking posted meetings (default: posted.jsonl)

Project Structure

city2facebook/
  main.py            # CLI entry point, argument parsing, workflow orchestration
  meeting.py         # Meeting data loading from archive bags
  extract_agenda.py  # PDF text extraction (PyMuPDF)
  summarizer.py      # LLM prompt building, API call, JSON parsing
  poster.py          # Post formatting, Facebook Graph API, duplicate tracking
  config.example.json # Template config (copy to config.json)
  requirements.txt   # Python dependencies

Performance Notes

• LLM summarization typically takes 3–5 minutes per meeting due to full transcript + agenda context
• Verbose mode (-v) logs token counts and generation time
• The list command uses optimized directory scanning to handle large archives (1,900+ meetings)