FASTPIX VIDEO SEARCH AGENT
Add "ask your video library" search to your product.
Multimodal video catalog search API with timestamped clip citations.
Your users ask a question in plain language and get a direct answer, with the exact clips cited at the right timestamp, across your whole VOD and live archive.
It is a search engine you embed in your product, not a tool your users leave to go use. Title-search returns the hour-long video; Video Search returns the 18-second moment that answers the question.
TRUSTED BY PRODUCT TEAMS SHIPPING VIDEO AT SCALE
What your users can do
Drop Video Search in, and your users can ask your library anything
Once Video Search is in your product, your audience asks in plain language and gets the exact moment back, with the cited clip. A few ways teams put it to work:
Connected fitness
"Get me 10 hamstring exercises for today."
Mid-class, a member asks for something specific. Video Search pulls the exact moments from your class library and queues them up, without anyone leaving the workout.
Newsroom
"Find the clip where this politician spoke on this."
Mid-interview, a producer asks the archive a question. Video Search returns the exact cited moment from years of footage, ready to roll on air in seconds.
Education
"Show me how this concept was taught."
A learner asks; Video Search returns the exact teaching moments. Assemble them into a custom lesson, for modern, ask-anything e-learning on your own catalog.
WHAT IT LOOKS LIKE IN YOUR PRODUCT
Query in. Answer out.
With cited moments.
Two pipes: indexing runs once per video on ingest. Query runs on every user search. Both share the same multimodal index.
frames + objects
built
query
question
ranked
citations
Sub-second on the query path for typical catalogs. Indexing is one-time on ingest; lives alongside the VOD asset.
Why catalog search is broken today
Your viewers ask questions. They get title-search results.
If your catalog has more than 500 videos, title-search is the wrong surface. Your viewers are typing questions; your UI is matching keywords.
The right moment is buried inside an hour-long video.
Title-search returns the lecture. The learner still has to scrub. The first three minutes don't answer the question. Drop-off compounds.
200 hours of manual tagging to make the catalog searchable.
Content ops teams hand-tagging topics, time-codes, and themes is the most-hated workflow in EdTech and media. It scales linearly with catalog. Then breaks when content changes.
You wrote "intelligent search" in your product brief 8 months ago.
Then realized building it means an embedding model, a vector store, a re-ranker, transcript chunking strategy, frame extraction, query expansion, a UX shell, evals, and a feedback loop. The shipping date moved twice.
CITED. FIVE DIMENSIONS PER ANSWER
Every cited clip is scored on five axes
CITED is how Video Search Agent decides which clips answer the question and which ones are background. Per-vertical defaults shipped. Per-customer tuning supported.
C
Confidence
How sure is the model that this clip answers the question? Score 0-1. Drives inline-vs-related routing in your UI.
I
Index match
Multimodal signal strength: transcript + frame + object + chapter score, jointly weighted.
T
Timestamp precision
Exact start/end of the answering moment. Not the whole video. The 18-second moment where the question is actually addressed.
E
Excerpt quality
The surfaced transcript chunk that proves the moment answers the question. Renders as a quote next to the clip.
D
Depth
Is the clip self-contained? Or does it depend on context viewers don't have? Filters out cliffhanger moments that need 90 seconds of setup.
EVERY QUERY GETS A CITED ANSWER
100% routed by your UX. Not 60/40.
Confidence ≥0.8 surfaces inline as the primary answer. Confidence 0.5-0.8 surfaces as "related moments." Confidence under 0.5 surfaces as a fallback search across the catalog. Three lanes; same Agent.
FIVE STEPS, START TO ANSWER
Index. Query. Score. Answer. Cite.
01
Index
Catalog ingests once. Multimodal index built across transcripts, frames, objects, chapters. Lives with the asset.
02
Query
User types or speaks a question. Query expanded; intent classified.
03
Score
Every candidate clip scored on CITED. Top-N retained; threshold lanes computed.
04
Answer
Structured answer composed from top clips. Sections, bullets, supporting quotes assembled.
05
Cite
Each section returns: clip + start/end timestamp + excerpt + thumbnail + playback_id.
WHAT MAKES IT POWERFUL
Search that understands video, not just titles.
Multimodal, not text-only
The index reads transcripts, scene bondries, on-screen frames, objects or face detection, OCR, NER and audio events, so a question resolves to the right moment even when the title says nothing.
The exact moment, cited
Every answer comes with the clip, the start and end timestamp, a quotable excerpt, a thumbnail, and a playback ID, ready to render and play.
Sub-second answers
Indexing is one-time on ingest, so the query path returns in under a second for typical catalogs. No re-build per query.
Follow-up questions
A follow-up token keeps conversation state, so viewers can refine and dig in without starting the search over.
Ask in any language
Indexing covers transcripts in multiple languages plus scene and object signals, so a viewer can ask in their own words and still land the cited moment.
One index, one stack
The same multimodal index powers In-Video AI summary and the Clipping Agent, lives next to the asset, and is priced as part of the stack, not a separate understanding bill.
PLUGS INTO YOUR STACK
One Agent. Every catalog. Every surface.
Video Search Agent runs on the same FastPix multimodal index that powers In-Video AI summary and Clipping Agent CHAIN scoring. The index is built once. Three products read from it.
VOD assets. uploaded videos
Every video uploaded to FastPix VOD is indexed automatically (opt-in). No re-encode required.
Live recordings. auto-archive
Live streams with auto-archive become VOD assets. Same index path. Live archive becomes searchable the moment recording completes.
External catalog. bring your own
If your video lives elsewhere (S3, Cloudflare R2, partner CDN), point the indexer at the source URL. Index runs without re-ingest.
FEATURED CUSTOMER
How Aadhan replaced title-search with question-answer + cited news moments.
Aadhan ships short-form news to a large multilingual audience. Their catalog of broadcast news segments was searchable by headline only. Viewers typed questions like "what did the minister say about inflation?"; the UI matched keywords in headlines, not what people said. The engineering team wired Video Search Agent against the existing FastPix Live archive: multimodal index across multiple languages, dropped a query UI into the Aadhan app, shipped behind a feature flag.
Read the Aadhan case study
Buyer journey
How Video Search Agent shows up across FastPix.
Four cluster pages that go deeper than this one. Each is a self-contained landing for a specific buyer journey.
AI video search across catalog
The shipping-customer pattern from Aadhan. Multimodal index, news-shaped queries, cited 30-second moments. Architecture that powered it.
Read the use caseLecture to study shorts
Long lectures become topic-tagged study shorts with AI-search across the catalog. EdTech platform pattern.
Read the use caseIn-Video AI deep dive
The multimodal indexing layer that powers Video Search Agent, Clipping Agent CHAIN scoring, In-Video AI summary, and auto-tagging.
Read the product pageHow Aadhan built news-catalog Q&A
Engineering team at a short-form news platform. Multimodal index across multiple languages of broadcast archives. Viewers ask; Agent answers with cited moments.
Read the storySEARCH AGENT FAQ
Questions buyers ask before going to production
What is multimodal indexing?
Transcripts plus scene detection plus object/face recognition plus caption/chapter markers, indexed jointly. The Agent answers across all four signals, not just title or description metadata. That's what makes "show me the moment Professor Singh factors a quadratic" resolve correctly.How fast is the answer?
Sub-second on the query path for catalogs under ~10K assets. Indexing cost is one-time per video on ingest. Lives alongside the VOD asset; no re-build on every query.Can it answer follow-up questions?
Yes. The Agent maintains conversation state via the follow_up_token in the response. Pass that token into the next query. Follow-ups are scoped to the prior answer and the same catalog index.Does it work for live archives?
Yes. Live recordings with auto-archive flow into the same multimodal index as VOD. The Agent answers across both.How does it compare to Algolia or Elastic?
Algolia and Elastic index text. Video Search Agent indexes video - transcripts plus frames plus objects plus chapters, jointly weighted. Different category. Most customers keep Algolia for text-content search and use Video Search Agent for video-content search.
Three ways to get unstuck
Whatever kind of help you need, there is a path.
Engineering support
Talk to a video engineer.
Stuck on an API call, a webhook signature, or a player integration? Reach the engineering team directly. Response within hours, not days.
Contact engineeringIntegration help
Docs, code samples, video tutorials.
Self-serve resources for the most common integrations. Quickstart guides, SDK examples, recorded walkthroughs on YouTube, and detailed playback logs in your dashboard.
Browse the docsSolution architect
Plan the rollout with a human.
New integration, migration off another platform, or a complex multi-tenant build. Book a session with a FastPix solution architect or join the Slack community to compare notes with other teams.
Join the Slack community