Spaces:
Sleeping
Sleeping
prianshujha commited on
Commit ·
3e3df89
1
Parent(s): 3e853a9
added readme.md
Browse files
readme.md
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# TeeleAgentHF
|
| 2 |
+
|
| 3 |
+
TeeleAgentHF is an AI-powered telecalling agent built for a Hugging Face competition. It captures live audio, transcribes speech, extracts scheduling intent, evaluates feasibility, and confirms bookings. Designed for low-VRAM deployment (4GB budget) and Hugging Face Spaces.
|
| 4 |
+
|
| 5 |
+
## Key Features
|
| 6 |
+
|
| 7 |
+
- Real-time microphone capture with Gradio UI
|
| 8 |
+
- ASR: Cohere Transcribe (streaming)
|
| 9 |
+
- Intent parsing: Qwen2.5-7B-Instruct (GGUF via llama-cpp-python)
|
| 10 |
+
- Evaluation: MiniCPM3-4B (int4 quantized evaluator)
|
| 11 |
+
- VAD: Silero VAD (ONNX)
|
| 12 |
+
- Persistent bookings in SQLite (`data/calls.db`)
|
| 13 |
+
- Scheduling rules and slot-checking logic
|
| 14 |
+
|
| 15 |
+
## Architecture
|
| 16 |
+
|
| 17 |
+
- app.py: Gradio front-end and session controls
|
| 18 |
+
- pipeline/: transcriber, intent parser, evaluator, orchestrator, VAD listener
|
| 19 |
+
- config.py & hf_config.json: model and inference configuration
|
| 20 |
+
- data/calls.db and db.py: call logging and booking persistence
|
| 21 |
+
|
| 22 |
+
## Requirements
|
| 23 |
+
|
| 24 |
+
- Python 3.10+ (3.11 recommended)
|
| 25 |
+
- CUDA-capable GPU for llama-cpp-python Qwen inference (recommended)
|
| 26 |
+
- Install dependencies: `pip install -r requirements.txt`
|
| 27 |
+
- Note: `llama-cpp-python` may require a CUDA-enabled build. Example:
|
| 28 |
+
```bash
|
| 29 |
+
CMAKE_ARGS="-DGGML_CUDA=on -DGGML_CUBLAS=on" pip install -U "llama-cpp-python"
|
| 30 |
+
```
|
| 31 |
+
|
| 32 |
+
## Running Locally
|
| 33 |
+
|
| 34 |
+
1. Create and activate a virtual environment
|
| 35 |
+
2. Install dependencies: `pip install -r requirements.txt`
|
| 36 |
+
3. Ensure models referenced in `hf_config.json` are available or accessible via Hugging Face
|
| 37 |
+
4. Start the app:
|
| 38 |
+
```bash
|
| 39 |
+
python app.py
|
| 40 |
+
```
|
| 41 |
+
5. Open http://127.0.0.1:7860 in a browser
|
| 42 |
+
|
| 43 |
+
## Deployment (Hugging Face Spaces)
|
| 44 |
+
|
| 45 |
+
- Ensure `app.py` listens on 0.0.0.0:7860 (config.py already uses these defaults)
|
| 46 |
+
- Provide model files or configure download/autoload in `hf_config.json`
|
| 47 |
+
- Verify VRAM budget and use quantized GGUF models to fit resource limits
|
| 48 |
+
|
| 49 |
+
## Configuration
|
| 50 |
+
|
| 51 |
+
- Edit `config.py` and `hf_config.json` to tune models, quantization, batch sizes, and scheduling rules (working hours, slot lengths, etc.)
|
| 52 |
+
|
| 53 |
+
## Collaborators
|
| 54 |
+
|
| 55 |
+
- Saurav Kumar Yadav <sauravkumaryadav100@gmail.com>
|
| 56 |
+
|
| 57 |
+
## Contributing
|
| 58 |
+
|
| 59 |
+
- Open issues or PRs. For large model changes, include resource and runtime notes.
|
| 60 |
+
|
| 61 |
+
## License
|
| 62 |
+
|
| 63 |
+
See LICENSE in the repository root.
|
| 64 |
+
|
| 65 |
+
## Contact
|
| 66 |
+
|
| 67 |
+
For questions about this project, contact the repository owner or listed collaborators.
|