One call from video to assistant
POST a video URL. The video is transcribed, its scenes are described, a knowledge base is indexed and an assistant is created — no collections or data sources to manage. Poll one status field until it says ready.
CatalogLatency Build: roughly the length of the video. Chat: first tokens in 1–3 seconds when streaming. Async true Rate Limit Per API key Max Input One video per video chat, reachable by public URL (MP4, MOV, WebM, MKV).
cn8/video-chat
Send one video and get an assistant that knows it: what is said, what is on
Learn more about Video Chat Chat & RAGsyncInstant previewchatvideorag
- 1Video
- 2Analyze
- 3Chat
Try it
Video
Analyzed once. Chatting afterwards costs 2 credits per 1K tokens; setup is free. We keep your video for 7 days so it can play next to the chat — the assistant stays after that.
What you get
An assistant that has heard and watched your video. Every answer comes with the moments it is based on — click one and the player jumps there.
How do I invite my team?
Open Settings → Members and press Invite. You can paste several e-mail addresses at once.
The Members tab of the Settings page, Invite dialog open12:40
“You can paste a whole list of addresses here.”14:10
- Knows speech and what is on screen
- Suggested opening questions, in the video's language
- Same chat API for your own player or widget
Overview
Send one video and get an assistant that knows it: what is said, what is on screen, and when. Viewers ask questions in plain language and every answer comes back with timestamped citations, so your player can jump straight to the moment being talked about. One call builds everything — analysis, knowledge base and assistant — and you chat through the same endpoint as any other CN8 chatbot, from your backend or key-less from your own web page.
Key capabilities
Answers that point into the video
Every answer carries citations with start_ms, end_ms and a ready-to-use seek_ms. Citations come from the retrieved passages, not from the model's wording, so they always point at a real part of the video.
Hears and sees
The assistant knows the speech and the visuals: on-screen text, slides, objects and actions are captioned per scene. Silent videos and screen recordings work too.
Embed without exposing a key
Every video chat can be chatted with key-less from any web page — built for player-side widgets. Restrict it to your own domains with public_domains when you need to.
Opening questions included
Two lists come with every video chat, in the video's language: starter_questions that fit any video (summarize, key takeaways) and suggested_questions written from this video's content. Show them as chips over an empty chat.
When to use it
E-learning and training
Learners ask the lesson instead of scrubbing through it and land on the exact explanation.
Product demos and webinars
Let prospects ask a recorded demo about pricing, features or setup and jump to that part.
Support and how-to libraries
Put an assistant next to every tutorial video: "where do I set this up?" answers with the step and its timestamp.
Interactive video players
Drive chapter jumps, highlights and overlays from citations in your own player.
Input & output
input
Public URL of a video file, a name, optional language and assistant instructions
JSON body
output
A video chat with a chatbot_key; chat answers with timestamped citations (REST or SSE)
JSONServer-Sent Events
Guides & tips
How it works
- Create: the video is downloaded, speech is transcribed and every scene is described (on-screen text, objects, actions).
- Index: scenes and transcript are stored as time-bound passages in a knowledge base private to your account.
- Chat: each question retrieves the most relevant passages; the answer is written from them and the passages' time ranges come back as citations.
Wiring citations to your player
- On a citation click set
player.currentTime = seek_ms / 1000and play. - Show
textas the clickable line ("what happens there") with the time next to it;labelis the full range. - Use
start_ms–end_msto highlight the range on your timeline. - When streaming, render the text as it arrives and add the chips when the
citationsevent comes in.
Embedding on your site
- Key-less chat works from any site by default — no setup. To lock it to your pages, set
public_domains(e.g.["academy.example.com"]; subdomains included). - From those pages call
POST /v1/proxy/core-chatwithout an API key; the request's Origin/Referer is checked against the list. - Never ship your API key to the browser — call key-less (optionally domain-restricted) or proxy through your backend.
- Optional: send
playback_position_ms(the player's current time in milliseconds) with each core-chat question. Viewers can then ask about the moment they are watching ("what does this mean?", "what did he just say?", "what is written on the board?") without describing it. - Optional: when a viewer taps a suggested question, send
origin: "suggested"in the core-chat body. Statistics already recognise suggested questions by their text (light edits included); the flag only makes heavily rewritten ones count too.
Usage numbers
core-conversations/{chatbot_key}?from=&to=pages the threads of a period;pagination.totalis the count for the same filter.core-conversation-stats/{chatbot_key}?from=&to=gives the screen numbers: conversations, questions, the suggested questions asked most, languages, most cited chapters. Both take epoch ms or ISO 8601 and default to all time.
Getting good answers
- Set
languagewhen you know it; auto-detect can stumble on videos that open with music or silence. - Use
system_messageto set tone, scope and answer language ("Answer in Turkish, keep it under three sentences"). - Replace
suggested_questionswith your own if you want to steer what viewers ask first;starter_questionsstay generic and follow the video's language.
Specs
Schema
Request body
chatbot_idrequiredstringThe chatbot_key of the video chat.
messagerequiredstringThe viewer's question.
conversation_keystringReturned by the first answer; send it back to continue the same conversation.
streamboolean · default: falsetrue streams the answer as Server-Sent Events; citations arrive as a separate event after the text.
Response
data.responsestringThe answer text.
data.citationsarrayThe parts of the video the answer is based on, in playback order (up to 4). Only passages that support the answer are cited. Absent when nothing time-bound was used.
start_msintegerStart of the cited part, in milliseconds.
end_msintegerEnd of the cited part, in milliseconds.
seek_msintegerWhere to move the player: the moment mentioned in the answer when there is one, otherwise start_ms.
labelstringHuman-readable range such as "12:32–13:01".
textstringOne line saying what happens in that part — what is on screen, or the quoted speech. Made to be shown as the clickable citation.
kindstringscene (visuals + speech), transcript (speech) or section (a chapter of the video).
previewstringShort excerpt of the cited passage.
video_urlstringPlayback URL of the video.
Pricing
You pay for analyzing the video once and for the tokens of each conversation. Setting up and managing video chats is free.
| Service | Unit | Price |
|---|---|---|
| Video analysis (on create) | second of video | 0.05 credits/second |
| Chat | 1,000 tokens | 0.3 credits |
| Details, list, update, delete | request | Free |
A 10-minute video costs about 120 credits to set up, once.
The analysis cost is reserved when you create the video chat and settled to the real video length when it finishes; failed analyses are released in full.
Key-less chats (from any site, or only your public_domains if you set them) are billed to the account that owns the video chat.