Optimize context

The question your model will answer, usually the user's message. Candidates are scored against it. Up to 8,000 characters.

The context items to choose from, up to 500. See Candidates.

A stable name for the call site, such as support-assistant. Groups requests in the dashboard, selects the agent's budget, and scopes the quality check's tuning.

The model id you send this context to, like gpt-6.1-sol or claude-sonnet-5-5. Costs in the report are priced for it. Unknown models fall back to your project's default model; report.model shows which was used.

Turns on cache-safe mode: items seen earlier in this conversation keep their first decision. 1 to 128 characters: letters, digits, _ - . :. See Conversations and prompt caching.

The content, up to 60,000 characters. HTML, JSON, logs and test output are cleaned automatically.

Your identifier, returned in context and report.items. Use your document or chunk id. Up to 80 characters.

A label for where it came from. Any name you choose works, such as rag, web, history or your own intern_data. Used to break savings down by source and to filter requests in the dashboard. It doesn't change scoring.

Stored lowercase, up to 40 characters. Characters other than letters, digits, _, - and spaces become -, so Intern.Data is stored as intern-data.

The most tokens of context that can come back, from 200 to 200,000 (values outside are clamped). Strongest items are kept first. Defaults to the agent's budget set in the dashboard, otherwise the project default (6,000).

conservative, balanced or aggressive: the base relevance threshold (0.9, 1.4, 1.9 on the 0 to 3 scale). See Selection.

2 to 10 model ids you approve, from any providers. Each must be in the catalog or priced in Settings → Models, otherwise the request returns 422.

Your model's instructions. Helps judge how demanding the task is.

Your typical answer length, to price output in the switch estimate.

Your full prompt size including history, for example the last usage.input_tokens. Defaults to what Contextwaise sees.

The request id, as shown in the dashboard (req_…).

The kept candidates as { id, text, source }, in the order you sent them. Send these to your model and store them in the conversation. text is the cleaned version.

Every figure and decision behind the result, described below.

before, after, saved and savedPct: input tokens of the context as received and as returned.

before and after at the model's input price, scorer (the cost of relevance scoring) and avoided (what you save, net of scoring). In USD.

The model costs were priced for. null when your project's custom price was used.

The input price used, in USD per 1M tokens.

One entry per candidate. See Items.

The settings applied. threshold includes the agent's learned adjustment.

With a conversation id: { id, reused, protected }, the number of decisions reused, and whether they were safely stored.

Routing advice when models was sent. See Read the advice.

The quality check runs after the response, so it is null here. Its verdict appears on the request in the dashboard.

Total time in milliseconds from the request arriving, and the split by step: auth, project, score, select.

The candidate's id and source.

Whether it is in context.

relevant (kept), low_relevance (below the threshold), duplicate (repeats a kept item, see duplicateOf) or over_budget (relevant, but the budget was full).

Relevance from 0 (unrelated) to 3 (essential).

The size after cleanup, and as received. They differ when cleanup removed noise.

What cleanup removed, such as converted HTML to text or dropped 12 empty JSON fields. Present only when the item was cleaned.

For duplicates: the id of the kept item it repeats.

In cache-safe mode: the decision was made earlier in the conversation.