--- title: NINA emoji: 🚀 colorFrom: gray colorTo: red sdk: gradio sdk_version: 6.14.0 python_version: '3.13' app_file: app.py pinned: false license: apache-2.0 short_description: Chat with a video, grounded only in the footage --- # NINA — Neural Information & Narrative Abstraction Upload a recording (or paste a link from a known video site) and chat about it. NINA answers using **only** the video: what is *said* (Whisper, two passes, with a reliability gate), what is *shown* (CLIP retrieval and a vision-language model that reads the sampled keyframes), speaker turns where available, and a whole-transcript digest for summaries. The app that drives this Space is **https://app.nnlabs.pl/nina**. ## Running this Space * **Secrets:** `HF_TOKEN` (inference providers; the pyannote models are gated, accept their terms on this token) and `NINA_SIGNING_SECRET` (shared with the website's Worker; without it every request is refused). Optional: `YT_DLP_COOKIES_B64`, `YT_DLP_PROXY` for pasted links; `TOKEN_REQUIRE_SCOPE=1` once the Worker emits scoped tokens. * **Hardware:** CPU (`cpu-upgrade`); sleeps after 15 minutes and takes about two minutes to wake (Whisper, CLIP and the cross-encoder load at start). * **Limits:** 500 MB and 45 minutes per video; uploads are purged after 15 minutes. ## API Every endpoint takes a signed, single-use token minted by the website for that endpoint's scope. There is no anonymous access. | api_name | inputs | output | | --- | --- | --- | | `chat` | video (optional), payload JSON `{messages, transcript, force_analyze, force_answer, has_video}`, token | streams `[answer, analyzed, transcript]` | | `route` | payload JSON, token | `analyze_first` / `confirm` / `answer` | | `fetch` | url, token | `[video, title_or_error]` | | `run` | video, prompt, token | streamed answer (legacy single shot) | Refusals and failures start with `⛔`; the website shows them as errors and does not store them as answers. Config reference: https://huggingface.co/docs/hub/spaces-config-reference