> ## Documentation Index
> Fetch the complete documentation index at: https://docs.memories.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# FAQ

> Common questions about video sources, processing time, throughput, formats, and limits.

Quick answers to the questions we hear most. For exact error/limit tables see [Errors & Rate Limits](/reference/errors-and-rate-limits); for the walkthrough see the [Quickstart](/datalake/quickstart).

## Video sources & formats

<AccordionGroup>
  <Accordion title="What video sources can I ingest?" icon="video">
    Three ways, all via [`POST /videos`](/datalake/videos/upload-video-url) (or the file / resumable variants):

    * **By URL** — a direct, publicly-fetchable **HTTPS media URL**.
    * **By file** — upload the file directly; use the **resumable** flow for large files.
    * **Live streams** — attach an `rtsp` / `rtmp` / `http` **pull** source with [Open Stream](/datalake/streams/open-stream).
  </Accordion>

  <Accordion title="Which containers / formats are supported?" icon="file-play">
    `mp4`, `m4v`, `mov`, `mkv`, `webm`, `avi`, `flv`, `f4v`, `wmv`, `mpg`, `mpeg`, `3gp`, `ts`, `m2ts`. A non-video file returns **`415 unsupported_media`**.
  </Accordion>

  <Accordion title="Can I ingest videos from TikTok, YouTube, Instagram, or X?" icon="share-2">
    You provide each video as a **direct, fetchable media URL** (the actual video file) or upload the file yourself. We don't pull from a TikTok / YouTube / Instagram / X **page or account link** — resolve it to the underlying media URL first, then pass that as `source_url`. A platform page URL will fail with `source_unresolved`. Each video is ingested independently, so a batch is just many `POST /videos` calls.
  </Accordion>
</AccordionGroup>

## Processing time & throughput

<AccordionGroup>
  <Accordion title="How long does it take to index a video?" icon="clock">
    Indexing runs `preprocess → index → derive` and scales with the **content processed = duration × fps** (default `fps: 1.0`, i.e. \~1 frame/second). Short clips finish quickly; long video takes proportionally longer. Watch it live by polling [`GET /operations/{id}`](/datalake/operations/get-operation) — the `progress` field reports each stage and a percentage.
  </Accordion>

  <Accordion title="How many videos can be processed at the same time?" icon="layers">
    Videos index **in parallel**. Submit them all up front — each returns its own `operation` immediately (`202`) and they process concurrently, up to the concurrency allowed on your plan (a rolling 2-hour concurrent-indexing cap). You don't wait for one to finish before starting the next.
  </Accordion>

  <Accordion title="How do I estimate time for a batch — e.g. 50 clips of 30s?" icon="calculator">
    Because processing is parallel, a batch does **not** take *N × single-video time*. A useful estimate:

    ```
    batch time ≈ ceil(N / C) × T
      N = number of videos
      C = concurrent indexing slots on your plan
      T = per-video index time (scales with duration × fps)
    ```

    So 50 short clips complete in a handful of parallel waves, not 50 sequential runs. For a guaranteed **volume throughput / SLA**, [contact us](mailto:sales@memories.ai) with your expected daily volume and clip length and we'll size it.
  </Accordion>

  <Accordion title="Does fps change how long it takes (and what it costs)?" icon="film">
    Yes — indexing works on **frames = duration × `fps`**, so `fps` is your main speed/cost lever. Lower `fps` (e.g. `0.5`) indexes **faster and cheaper** with coarser temporal detail; raise it for fast action where you need finer moments. Default is `1.0`; range `0.1–30`.
  </Accordion>

  <Accordion title="Do safety detection or face recognition add time?" icon="shield-half">
    A little. They're **opt-in** collection flags that run extra derivation per video, so a collection with them enabled takes somewhat longer than a plain index. Leave them off on collections that don't need them.
  </Accordion>

  <Accordion title="What if I submit more videos than my concurrency allows?" icon="list-ordered">
    Each `POST /videos` still returns immediately, but ingests beyond your plan's concurrent-indexing cap come back **`429 rate_limited`** with a `Retry-After`. Back off per that header (or throttle client-side) and the queue drains as slots free up — nothing is lost.
  </Accordion>

  <Accordion title="When can I start searching a video — before it's fully done?" icon="hourglass">
    A video becomes searchable once its operation reaches **`ready`**. While it's still processing you can pull derived content **incrementally** with `?since=` — the response's `live: true` means indexing is still running, so keep paging until it clears.
  </Accordion>

  <Accordion title="How fast are results on a live stream?" icon="radio-tower">
    Live captioning and events are **near-real-time** — they surface as the stream plays, rather than after a whole file is uploaded. Latency tracks your `fps` and any enabled detectors. There's no fixed session-duration cap.
  </Accordion>

  <Accordion title="What happens if a video fails or times out?" icon="triangle-alert">
    The operation ends with `done: true` and a non-null `error` (e.g. `source_unresolved` for an unreachable link, `unsupported_media` for a bad format). Fix the source and re-submit — reuse the same `idempotency_key` so a retry never double-charges or double-ingests.
  </Accordion>
</AccordionGroup>

## Results & notifications

<AccordionGroup>
  <Accordion title="How do I know when a video is ready?" icon="bell">
    Every slow operation returns `202` + an operation. Either **poll** [`GET /operations/{id}`](/datalake/operations/get-operation) until `done: true`, or attach a **[webhook](/datalake/reference/webhooks)** (`callback_url`) and get a signed POST on completion — no polling. `done` is the only completion signal.
  </Accordion>

  <Accordion title="What can I read back from an indexed video?" icon="list">
    Per moment or whole video: **captions, transcripts, frames, clips, titles, summaries, speakers, face entities**, and **safety events** (if a detector is enabled). Read any time-slice with [Get Moment](/datalake/moments/get-moment), or search across a collection with [Search](/datalake/search/search).
  </Accordion>

  <Accordion title="What languages are supported?" icon="languages">
    Search and retrieval run on **OmniRetriever**, a multilingual embedding model — you can query in one language and match content in another.
  </Accordion>
</AccordionGroup>

## Billing & limits

<AccordionGroup>
  <Accordion title="Where do I see usage and billing?" icon="credit-card">
    In the **[Console](https://console.memories.ai)** → **Usage & billing** — balance, usage, and invoices. Your live balance is also available at [`GET /usage/balance`](/datalake/operations/get-balance).
  </Accordion>

  <Accordion title="What are the main limits?" icon="gauge">
    `fps` 0.1–30 (default 1.0) · list `limit` 1–100 · **Search 5 QPS/user**, `top_k` ≤ 200 · search image ≤ 10 MB · person reference faces 1–20/call, ≤15 MB each. Full table in [Errors & Rate Limits](/reference/errors-and-rate-limits).
  </Accordion>

  <Accordion title="How long do returned media URLs stay valid?" icon="link">
    Signed URLs expire — clip cache \~5h, detail `source_url` / face image \~24h, search thumbnails \~15 min. Re-fetch on demand; never cache past expiry.
  </Accordion>
</AccordionGroup>

<Note>
  Don't see your question? [Talk to us](mailto:support@memories.ai) or point your agent at the [MCP Server](/mcp-server) and just ask.
</Note>