CS2 RAG Performance Coach

Analyze your CS2 match demos with a fully local, zero-cost RAG pipeline. Ask natural language questions and get actionable coaching insights backed by your actual match data.

---

Prerequisites

Tool	Install
Python 3.11+	https://python.org
Ollama	https://ollama.com

---

1 — Install Python dependencies

pip install -r requirements.txt

---

2 — Pull Ollama models

# Embedding model (small, fast)
ollama pull nomic-embed-text

# LLM — pick ONE based on your RAM:
ollama pull llama3.2          # 8 GB RAM  (3B, fastest)
ollama pull mistral           # 16 GB RAM (7B, better reasoning)
ollama pull llama3.1:8b       # 16 GB RAM (8B, good balance)

---

3 — Download demos from Leetify

# Download 10 most recent demos (prompts for your Steam64 ID)
python fetcher/fetch_demos.py --steam-id 76561198012345678 --count 10

# With a Leetify API key (higher rate limits) — get one at leetify.com/app/developer
python fetcher/fetch_demos.py --steam-id 76561198012345678 --count 10 --api-key YOUR_KEY

# Or set the key as an env var so you don't have to type it each time
export LEETIFY_API_KEY=your_key_here
python fetcher/fetch_demos.py --steam-id 76561198012345678 --count 10

# Force re-download demos already present
python fetcher/fetch_demos.py --steam-id 76561198012345678 --count 10 --redownload

Demos are saved to data/demos/ as <game_id>.dem, decompressed automatically.

---

4 — Parse a demo

Drop your .dem files into data/demos/, then run:

python parser/parse_demo.py data/demos/your_match.dem

Parsed round chunks land in data/parsed/your_match.json.

---

4 — Ingest into ChromaDB

# Ingest all parsed files
python ingest/embed_and_store.py

# Or a specific file
python ingest/embed_and_store.py --file data/parsed/your_match.json

---

5 — Ask questions

CLI:

python retrieval/query.py "Why do I keep losing pistol rounds?"

# Filter to a specific map
python retrieval/query.py "How is my B site defence?" --map de_mirage

Web UI:

python ui/app.py
# Open http://localhost:7860

---

Project Layout

cs2-rag/
├── fetcher/
│   └── fetch_demos.py       # Leetify API → download .dem files
├── parser/
│   └── parse_demo.py        # awpy demo → round chunks (JSON)
├── ingest/
│   └── embed_and_store.py   # embed chunks → ChromaDB
├── retrieval/
│   └── query.py             # similarity search + LLM generation
├── ui/
│   └── app.py               # Gradio web interface
├── data/
│   ├── demos/               # place .dem files here
│   ├── parsed/              # auto-generated JSON chunks
│   └── chroma_db/           # auto-generated vector DB
└── requirements.txt

---

Example Questions

• "Why am I losing pistol rounds?"
• "Which map is my worst?"
• "Am I making good economic decisions?"
• "What weapons kill me most?"
• "How often do I win after planting the bomb?"
• "When should I be saving instead of force-buying?"

---

Tuning Tips

• More context: Increase TOP_K in query.py if answers feel too narrow.
• Better LLM: Switch LLM_MODEL to mistral or llama3.1:8b for deeper analysis.
• Map filters: Use --map de_inferno to focus retrieval on a single map.
• Re-ingest is safe: embed_and_store.py uses upsert, so re-running won't duplicate.