Add AI chat demo with prompt-based embedding ad targeting

[ericholscher]

Summary

Prototype of an AI chat interface that targets ethical ads based on conversation context using embeddings.

• Chat UI at /chat/ sends user prompts to OpenAI (gpt-4o-mini) for responses
• In parallel, the same prompt is sent to the ad decision API with a new prompt field
• The backend generates an embedding (OpenAI text-embedding-3-small) from the prompt text and computes cosine similarity against advertiser content embeddings
• Flights with niche_targeting set in their targeting_parameters are filtered by this similarity score
• When no prompt is provided, the existing URL-based niche targeting via ethicalads_ext.embedding is used as a fallback
• Keywords are also extracted client-side from the prompt as a secondary targeting signal

Targeting approach

1. Prompt embedding: The full prompt text is embedded via OpenAI's text-embedding-3-small model and cached for 1 hour
2. Advertiser embedding: Each flight's live ad creatives (headline, content, CTA) are embedded the same way
3. Cosine similarity: The distance between prompt and advertiser embeddings is compared against each flight's niche_targeting threshold (lower distance = better match)
4. Keyword fallback: Significant words are extracted client-side from the prompt and passed as pipe-delimited keywords for standard keyword targeting

Local setup

# Install and migrate
pip install openai
python manage.py migrate

# Prep the database with test data (uses first existing objects, creates what's missing)
python manage.py shell < scripts/prep_chat_demo_db.py

# Set env vars (publisher slug is printed by the script above)
export OPENAI_API_KEY="sk-your-key-here"
export ADSERVER_CHAT_DEMO_PUBLISHER="your-publisher-slug"

# Run and open http://localhost:8000/chat/
python manage.py runserver

The prep script grabs the first existing Publisher, Advertiser, Campaign, Flight, Advertisement, and AdType, then patches them to be live and wired together for the demo.

Test plan

• Run python manage.py shell < scripts/prep_chat_demo_db.py on a fresh DB and verify it creates all needed objects
• Run on a DB with existing data and verify it reuses first objects without creating duplicates
• Open /chat/, send a message, verify chat response appears
• Verify ad decision request fires in the browser network tab with prompt and keywords params
• Verify ads render in the sidebar when matching ads exist
• Test with OPENAI_API_KEY unset: chat returns error, ad API still works via keyword targeting
• Test the decision API directly: curl 'http://localhost:8000/api/v1/decision/?publisher=SLUG&div_ids=ad1&ad_types=AD_TYPE&prompt=python+web+development'

https://claude.ai/code/session_01YGtppBeFRZNHxdEguYKxnY