Gemma 3 L&D Assistant POC

This repo holds a learning-and-development assistant prototype powered by Google’s Gemma 3 models. It ingests public financial statements, chunks them, embeds the text into a Chroma vector store, and retrieves grounded context for Gemma 3 to answer banker-style questions with citations. The L&D code now lives under ld/ to keep it separate from the skin-care playground tools.

Documentation

• Getting Started
• Ingestion & Chunking
• Vectorization
• Retrieval & Answering
• Roadmap

Use the docs for detailed walkthroughs; the sections below summarize the core pipeline.

1. Ingest PDFs with ld/ingest.py (writes JSONL chunks).
2. Vectorize chunks with ld/vectorize.py (SentenceTransformer → Chroma).
3. Query via ld/answer.py (single shot) or ld/chat_cli.py (interactive, with memory/logging).
4. Iterate using the roadmap checkpoints to add structured tools, recommendation logic, and guardrails.

Skin-care news digest (no RAG)

The skin-focused experiments now live under skin/. Use skin/skin_news.py when you need a real-time digest without persisting embeddings:

uv run python skin/skin_news.py \
  --source-file skin/samples/skin_sources_sample.json \
  --limit 3 \
  --cache-file data/cache/skin_news_cache.json \
  --ttl 900

• Reads recent source snippets (the sample file under skin/samples/ mocks YouTube transcripts; swap in real fetcher output).
• Builds a Gemma 3 prompt on the fly and returns a Markdown digest tuned for a fantasy “skin stock” mechanic.
• Saves the latest result under data/cache/... so repeated calls within the TTL reuse the summary.
• Pass --force-refresh to bypass the cache or --output digest.md to capture the rendered digest.

Live YouTube pulls

Point the same script at real channels by enabling the YouTube Data API (set YOUTUBE_API_KEY in .env):

uv run python skin/skin_news.py \
  --youtube-channel-id https://www.youtube.com/@MasterShiny \
  --youtube-channel-id https://www.youtube.com/@SkinomiCS2 \
  --youtube-playlist-id https://www.youtube.com/playlist?list=PLjRE1qx9Gn8jow1mHKumOlmV01Gtc0JCy \
  --youtube-max-results 3 \
  --youtube-playlist-max-results 2 \
  --skip-source-file \
  --limit 5

• Fetches the latest videos per channel, grabs transcripts via youtube-transcript-api, and feeds them straight to Gemma.
• Accepts channel IDs, handles (@channel), or full URLs; playlist URLs/IDs are supported via --youtube-playlist-id.
• Reuses the same cache (keyed by video IDs) so repeated calls within the TTL stay instant.
• Keep --source-file around for fixtures; add --skip-source-file (as above) to ignore it when hitting real data.

Find additional skin-specific notes in skin/README.md.