ppt-copilot

Production-ish applied-science demo for a PowerPoint copilot:

• ingest slides
• summarize
• retrieval-grounded QA with citations
• scenario eval harness
• finetune reproducibility pipeline
• latency/cost benchmarking

5–10 Minute Quickstart

1) Install

python -m pip install -e .

2) Create sample deck + ingest

python scripts/generate_sample_pptx.py -o data/raw/sample_deck.pptx
pptcopilot ingest data/raw/sample_deck.pptx -o data/processed/sample.jsonl

Expected output:

{"status":"ok",...,"n_records":6}

3) Index + summarize + QA

pptcopilot index data/processed/sample.jsonl --index-dir artifacts/index/sample
pptcopilot summarize data/processed/sample.jsonl --model mock -o reports/summaries/sample_summary.json
pptcopilot qa --index-dir artifacts/index/sample --question "Which slide discusses evaluation harness?" -o reports/qa/sample_qa.json

Expected artifacts:

• reports/summaries/sample_summary.json
• reports/qa/sample_qa.json

4) Run eval suite

pptcopilot eval run --suite smoke --model mock --outdir reports/eval

Expected artifacts:

• reports/eval/latest_report.json
• reports/eval/latest_report.md

5) Run benchmarks

pptcopilot bench latency --suite smoke --out reports/bench/latency.json --iterations 5 --warmup 1 --batch 1
pptcopilot bench cost --in reports/bench/latency.json --out reports/bench/cost.json --summary-out reports/bench/summary.md

Expected artifacts:

• reports/bench/latency.json
• reports/bench/cost.json
• reports/bench/summary.md

Optional API Demo (FastAPI)

Install API extras:

python -m pip install -e ".[api]"

Run server:

uvicorn pptcopilot.api.app:app --host 0.0.0.0 --port 8000

Endpoints:

• POST /summarize
• POST /qa
• POST /eval/run

Example:

curl -s http://localhost:8000/summarize \
  -H "Content-Type: application/json" \
  -d '{"input_jsonl":"data/processed/sample.jsonl","model":"mock"}'

CI

Workflow: .github/workflows/nightly_eval.yml

• PR/push: runs pytest
• nightly cron: runs smoke eval
• uploads reports/eval/latest_report.json + .md as artifacts

Finetune Reproducibility

Smoke finetune:

pptcopilot finetune --config configs/finetune_small.yaml

Details: reports/finetune/README.md.

Demo Capture Instructions

Terminal capture (portable):

script -q demo_terminal.txt
# run ingest -> index -> summarize -> qa -> eval commands
exit

Optional GIF route:

1. Record terminal with any recorder (asciinema rec demo.cast or screen recorder).
2. Convert to GIF (agg, asciinema + svg-term, or editor export).
3. Attach the GIF to your repo/PR.

Failure Handling

• CLI now emits structured JSON errors and logs stack traces.
• API endpoints return HTTP 400 with informative detail messages.