Add README with Hugging Face Space YAML configuration.

README.md

@@ -0,0 +1,59 @@
+---
+title: WitGym
+emoji: 🎭
+colorFrom: green
+colorTo: yellow
+sdk: gradio
+sdk_version: "5.0.0"
+app_file: app.py
+pinned: false
+license: mit
+tags:
+  - comedy
+  - rag
+  - case-based-reasoning
+  - qwen
+---
+
+# WitGym
+
+Conversational wit grounded in human comedy precedent — a Case-Based Reasoning RAG (CBR-RAG) engine.
+
+Paste a real-life awkward situation. WitGym extracts the comedy structure, retrieves analogous precedents, generates persona candidates, ranks the sharpest line, and shows every intermediate step in the logs panel.
+
+## Run locally
+
+```bash
+# Install
+uv sync  # or pip install -e .
+
+# Build index (The Office transcripts only)
+witgym-index
+
+# CLI (local 9B on MPS/CUDA)
+python -m witgym.main --debug
+
+# Gradio UI (local model)
+python app.py
+
+# Gradio UI via Hugging Face Inference Providers (no local 9B weights)
+export HF_TOKEN=hf_...
+export LLM_BACKEND=hf_api
+python app.py
+```
+
+## Hugging Face Space secrets
+
+| Secret | Value |
+|--------|--------|
+| `HF_TOKEN` | Your Hugging Face access token |
+| `LLM_BACKEND` | `hf_api` (recommended on Spaces) |
+
+Optional: `HF_INFERENCE_PROVIDER` (defaults to `together` for Qwen3.5-9B — required for `enable_thinking: false`), `WITGYM_INDEX_PATH`.
+
+## Architecture
+
+- **Small talk** (hi, who are you) → polite identity reply, no pipeline
+- **Humour practice** → Pass 1 metadata → RAG retrieval → Pass 2 personas → rank → compress
+
+Backend: `local` (Transformers) or `hf_api` (Inference Providers via `huggingface_hub.InferenceClient`).