Splyc AI draws on two complementary systems—a corpus (broad patterns from clips in production) and a live knowledge layer (what’s happening recently, plus published guidance). Together they keep answers useful and current without exposing private clip data in bulk.
Why two systems?
- Corpus answers: “What do successful clips tend to look like?”—titles, tags, phrasing by category, short example lines—built from large volumes of real production clips, refreshed on a schedule.
- Live knowledge answers: “What’s the product seeing lately?”—rolled-up, anonymized activity (for example, how much automated analysis and thumbs feedback is landing per category) plus responsible clipping policy text, updated as the product is used.
Chat does not rewrite those stores directly. They are maintained through logged analysis, scheduled jobs, and team-controlled policy, so behavior stays predictable and auditable.
The corpus (slow, wide loop)
Splyc compresses production clip data into reference packs: tag and title patterns (including by clip category), and short anonymized snippets that can be matched to a question or to your clip’s metadata.
How it stays fresh:
- When media analysis runs on clips, the product records structured summary metadata (for example, quality tiers, ratings, cache flags, category)—enough to improve retrieval and quality over time, not raw video.
- On a regular schedule, those records are consolidated per clip and merged with things like category and extension feedback where available.
- Reference packs are rebuilt from eligible clips (global and, where applicable, per organization), respecting content and visibility rules (for example, public-only vs broader modes, safety filters).
Splyc AI can sound more grounded in real stream culture because it is reading curated summaries of production behavior, not guessing from a generic model alone. The packs are versioned snapshots—the product reads the latest pack, not the entire clip database on every message. Separately, Splyc can include curated culture and ecosystem context so suggestions respect how given corners of the internet actually talk.
Live knowledge (fast, recent loop)
A rolling picture of recent product activity, bucketed by week and category, plus policy your team can publish (defaults exist; staff can override).
What feeds it:
- When an extension clip finishes automated media analysis, counters can advance for eligible clips (same high-level gates as global corpus policy: for example, visibility and safety).
- Thumbs up / down (and neutral) on AI-captured extension clips update feedback ratios in those same buckets.
What shows up in Splyc AI:
- Roughly the last several weeks of those rolled-up counts.
- Optional short trend summaries when the summarizer job is enabled.
- Responsible clipping guidance, so the assistant stays aligned with published norms.
You get recency—what the fleet is doing this month—without waiting for a full corpus rebuild. Chats read this layer; they do not append to it.
How suggestions are assembled
When you open the clip assistant or the floating agent with a clip, Splyc typically combines:
| Building block | What it is |
|---|---|
| Your clip | Transcript (within limits), metadata, timeline context—the primary source of truth for that session. |
| Corpus hints | Tag/title patterns from the latest reference packs (global and your org when applicable). |
| Matched examples | Short anonymized lines retrieved from those packs when they match your clip or question. |
| Live knowledge | Recent anonymized rollups + policy text. |
| Culture | Matched ecosystem/context blocks when enabled. |
| Voice | Streamer-specific phrasing notes when a voice profile exists. |
| Your settings | Video profile AI hints when you’ve turned them on. |
If you use the agent without an active clip, you still get corpus hints and live knowledge where enabled—no automatic voice/culture block tied to a specific moment. Running analysis in a session feeds future corpus updates on the next scheduled pass; it does not instantly rewrite the whole reference library mid-chat.
Privacy and safety
- Corpus snippets are built under published rules; prompts describe them as anonymized example material, not “here is user X’s private library.”
- Live knowledge stores aggregates (counts by week and category), not a public leaderboard of individuals.
- Internal usage logs may still tie events to accounts for billing and operations; what the model sees for retrieval is artifact-based, not a raw export of every clip row.
At a glance: how knowledge currently recycles
- Clips run through analysis → structured metadata is recorded; eligible activity bumps live knowledge counters.
- On a schedule → metadata is rolled up per clip, then reference packs can be rebuilt so patterns and snippets stay current.
- On each request → Splyc AI gets your clip, latest packs, live rollups + policy, and (when relevant) culture and voice.
No single chat message is the source of truth for the corpus; pipelines and policy are.