Summarizer model & background AI
Every provider in DeepState has two model slots: the main model the assistant talks with, and a summarizer model for the many small jobs that run in the background. Using a fast, cheap model for the second slot keeps the app responsive and your bill sane; the defaults do exactly that.
What the summarizer does
| Job | When |
|---|---|
| Article summaries - summary, key points, sentiment, topics on News Article nodes | Whenever an article is scraped (paste, feed, fetch_and_create_newsfeed, re-scrape) |
| Feed relevance filter | On every feed poll where the filter is enabled (RSS feeds) |
| Narrative suggestions, narrative summaries, relationship analysis | The right-click and context-panel AI actions (Narratives) |
| Conversation compaction | When you press Compact conversation in the agent's header |
| Autoresearch supervisor | Between rounds of an Autoresearch session |
| Long tool results - transcripts and SEC filings over a few thousand characters | Automatically, to keep them inside the assistant's context |
None of these need the main model's depth; all of them run often.
Defaults
| Provider | Summarizer default |
|---|---|
| DeepState AI | anthropic/claude-haiku-4.5 |
| Claude | claude-haiku-4-5 |
| OpenAI | gpt-5-mini |
| OpenRouter | anthropic/claude-haiku-4.5 |
| OpenCode Zen | anthropic/claude-haiku-4.5 |
| OpenCode Go | glm-5 |
| Ollama | the same model as your main model |
Change it under Settings → AI Provider → Summarizer model. The same model list is offered as for the main slot.
Choosing
- Keep it small. Haiku / mini-class models summarize an article in a second or two. The main model would produce a marginally nicer summary at ten times the cost, hundreds of times a day on a busy board.
- But not too small. The feed relevance filter and the Autoresearch supervisor need judgement. If feeds are letting junk through or the supervisor directives are vague, step the summarizer up a class before touching anything else.
- Local models. With Ollama, a 7-8B instruction model handles summaries well. Relevance scoring is more variable; test on a feed you know.
Feed relevance in detail
When AI relevance filter is on for a feed, each new item's title and description are sent to the summarizer with the board's name, description and a sketch of what is on it, plus your optional custom prompt ("we care about litigation and executive departures, not product news"). The model returns a relevance judgement with a confidence from 0 to 1; items at or above the feed's threshold (default 0.5) are added to the board, the rest are recorded as filtered so they are not re-evaluated. Keyword include/exclude filters run first and are free.
Article summaries in detail
The summarizer receives the extracted article text and returns a short summary, three to five
key points, a sentiment (positive / negative / neutral / mixed) and topic tags. These are stored
on the node, shown on the card, and searchable by the assistant's search_board. Very long
articles are truncated before summarization; the full text is still kept on the node.
Cost and budget
Summarizer traffic counts against your provider usage like anything else. On DeepState AI it draws from the same monthly allowance as the assistant. A board with a dozen active feeds and relevance filtering on can generate a few hundred small calls a day - cheap with a Haiku-class model, noticeable with a frontier one.