| --- |
| title: Final Assignment Agent |
| emoji: 🕵️ |
| colorFrom: indigo |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 6.29.1 |
| python_version: 3.10 |
| app_file: app.py |
| pinned: false |
| hf_oauth: true |
| hf_oauth_expiration_minutes: 480 |
| --- |
| |
| # Agents Course — Unit 4 |
|
|
| An evidence-based agent using Groq, smolagents tool schemas and a native agent loop. |
| Add `GROQ_API_KEY` as a Space Secret. The default text model is |
| `openai/gpt-oss-120b`. Optional variables: `GROQ_MODEL` and `GROQ_VISION_MODEL`. |
| The default vision model is `qwen/qwen3.8-27b`; audio uses `whisper-large-v3-turbo`. |
| The app does not change your Groq plan or call the Space GPU. Free-plan quotas apply. |
|
|
| Sign in as the Space owner. Run one question or the remaining questions, review |
| answers and tool traces, then use the separate official submission button. |
| Successful answers are saved in private runtime storage and can be restored |
| on reconnect. A Space rebuild/reset removes runtime files; JSON answer backups |
| can be imported through the owner-only restore controls. |
|
|
| Tools support web search, historical English Wikipedia revisions and sections, |
| paginated webpages/PDFs, page-text lookup, arithmetic, XLSX reading, restricted |
| Python interpretation, audio transcription, and image analysis. Chess positions |
| are checked with Stockfish. Unsupported Python constructs are returned as source |
| for reasoning, with an explicit execution limitation. Video scene analysis is |
| not implemented. Unknown answers and failed questions are omitted from submission. |
|
|
| After provider tool-parser errors, the agent retains observations and retries |
| with a JSON action planner. A valid rejected JSON tool call may be recovered only |
| when its name is allowlisted; failed-generation prose is never accepted as an answer. |
| Every draft answer gets a second pass to check source, constraints and format. |
| No benchmark answer keys, copied submission answers or hardcoded question answers |
| are included in the code. Optional source URLs can guide a retry to original evidence. |
|
|
| The official course attachment endpoint can return 404. Original input files may |
| be restored through the owner-only input control after accepting GAIA's access |
| conditions. Inputs are kept in private runtime storage, never in the public Git |
| repository or publicly served Gradio file cache. Only the owner may run tools, |
| restore data or submit. No API keys or provider reasoning are displayed. |
|
|
| ZeroGPU startup registers an uncalled compatibility function; model inference |
| runs on Groq. No benchmark score or certificate result is guaranteed. |
|
|
| References: |
| - [Official course](https://huggingface.co/learn/agents-course/unit4/hands-on) |
| - [Groq tool calling](https://console.groq.com/docs/tool-use/local-tool-calling) |
| - [Groq vision](https://console.groq.com/docs/vision) |
| - [Groq audio](https://console.groq.com/docs/speech-to-text) |
| - [GAIA input dataset and conditions](https://huggingface.co/datasets/gaia-benchmark/GAIA) |
|
|
