Deepgrip From $89/mo · Free plan 9.3 Visit MomentSearch
github.com
What it is
MomentSearch is open-source software for visual video search and retrieval-augmented answers tied to specific video moments. Users can upload videos or submit YouTube URLs for indexing. Each search checks both visual frames and transcript text, then combines the results. Answers can point to clickable moments with keyframes, timestamps, synchronized transcript text, labels showing whether a match was seen or said, and match scores. A confidence gate can return no generated answer without making an LLM call when neither search branch meets its threshold. Without an LLM key, the system still returns ranked clickable moments, but does not write prose answers. Answer, embedding, reranking, and speech-to-text models can be selected by provider, with hosted and local options documented. Optional speaker recognition uses Gemini video understanding and requires a Gemini API key. The project includes Docker entrypoints and deployment guidance for Fly, AWS, and Google Cloud. Its Apache 2.0 license permits open-source use. The API has no authentication, and the application is single-user by default.
Spec sheet
- Purpose
- MomentSearch lets users ask questions about videos and retrieves exact moments using what appears on screen and what is said in transcripts.Read 1 Oct 2026 · github.com
- Deployment
- MomentSearch is packaged as a Docker image with api, worker, clip, and seed entrypoints and can deploy anywhere Docker runs.Read 1 Oct 2026 · github.com
- Integrations
- Documented infrastructure integrations include Postgres, Qdrant, Prefect, and object storage through Tigris, Amazon S3, or Google Cloud Storage.Read 1 Oct 2026 · github.com
- License
- MomentSearch is released under the Apache 2.0 license.Read 1 Oct 2026 · github.com
- Abstention
- A confidence gate can abstain without making an LLM call when neither retrieval branch clears its threshold.Read 30 Sep 2026 · github.com
- Citations
- Answers include clickable citations with matched keyframes, timestamps, transcript synchronization, evidence badges, and match scores.Read 1 Oct 2026 · github.com
- Default models
- The documented default stack uses OpenAI GPT-4o for answers, OpenAI text-embedding-3-small for transcripts, local CLIP for frames, fastembed for reranking, Whisper-1 for speech-to-text, and Gemini for speaker recognition.Read 1 Oct 2026 · github.com
- Input
- Users can upload videos or paste YouTube URLs for indexing.Read 1 Oct 2026 · github.com
- Maker privacy
- Traversaal.ai says it uses strictly necessary authentication and usage cookies and does not use cookies for advertising, analytics, or tracking.Read 30 Sep 2026 · traversaal.ai
- Model flexibility
- Answer, embedding, reranking, and speech-to-text models are configurable by provider name, including hosted and local options.Read 1 Oct 2026 · github.com
- Models
- Answer, embedding, reranking, and speech-to-text models are pluggable by provider name, with OpenAI, Gemini, Anthropic, OpenRouter, xAI, Groq, Together, Fireworks, Mistral, NVIDIA, Azure OpenAI, Ollama, LM Studio, vLLM, Cohere, Voyage, Jina, and local CLIP options documented.Read 30 Sep 2026 · github.com
- Retrieval
- The system searches visual frames and transcripts in parallel, fuses the results, and reranks matching moments.Read 1 Oct 2026 · github.com
- Retrieval-only mode
- Without an LLM key, the application still returns ranked clickable moments but does not generate prose answers.Read 1 Oct 2026 · github.com
- Security limit
- The API is unauthenticated; anyone who can reach its port can mint upload URLs, ingest, ask, and delete, so the README recommends localhost binding or an authenticating proxy.Read 30 Sep 2026 · github.com
- Security limitation
- The API has no authentication, and anyone who can reach the port can upload, ingest, ask, and delete data.Read 1 Oct 2026 · github.com
- Speaker recognition
- Optional per-video speaker recognition uses Gemini video understanding to label who said each line.Read 1 Oct 2026 · github.com
- Storage integrations
- Object storage supports local storage, AWS S3, GCP storage, GCP native storage, and Fly Tigris, while the architecture uses Postgres, Qdrant, and Prefect Cloud.Read 30 Sep 2026 · github.com
- Storage protection
- The project recommends keeping buckets private and using presigned URLs for uploads and playback.Read 1 Oct 2026 · github.com
- Uploads
- Presigned uploads send gigabytes directly from the browser to object storage instead of through the API.Read 30 Sep 2026 · github.com
- User model
- The application is single-user by default and ignores X-User-Id and Authorization headers, routing every request through SINGLE_USER_ID.Read 30 Sep 2026 · github.com
- YouTube ingestion limit
- The documented SocialKit YouTube ingestion option provides 20 free credits and then becomes paid, while cookie authentication is free but cookies expire approximately every two to three weeks.Read 30 Sep 2026 · github.com
Strong points and limits
Strong points
- Search combines visual frames with transcript text
- Citations include keyframes, timestamps, and match scores
- Models can use hosted or local providers
- Apache 2.0 licensed
Limits
- API has no authentication
- Application is single-user by default
- Speaker recognition requires a Gemini API key
- SocialKit YouTube ingestion becomes paid after 20 free credits
Who it suits
MomentSearch suits developers building or self-hosting video search systems who need both visual and transcript retrieval. It also fits users who want timestamped, clickable results rather than prose answers alone.
Verdict
MomentSearch offers multimodal retrieval with clickable, timestamped results and configurable models. Its unauthenticated API and single-user default require careful deployment choices.
Where Specifiction places it
- AI Video Search Tools#15 of 247.6
