Long-context analysis
Keep large inputs in variable space. Inputs may come from the user query, committed Session History, Skill resources, Attachments, a public URL, or Session Workspace. Explore them with bounded Python operations, send only relevant excerpts to sub-model calls, and verify every reported fact against the original variable before submitting.
Analyze
- Inspect variable names, types, lengths, and small previews. Do not print a whole large value. For a relevant URL, call
fetch_url once and keep its returned content in a variable.
- Locate candidate regions with deterministic searches, indexes, regular expressions, or bounded slices.
- When the relevant regions are unknown, scan every bounded chunk for structured candidates before reducing them. Query ranking may prioritize reading order, but it must not exclude unseen evidence.
- Call
llm_query or llm_query_batched only on self-contained excerpts that include the question and their source offsets or source identifiers. Use rlm_query only for the rare selected subproblem that needs a fresh iterative Python investigation; it is not the normal route for extraction, counting, parsing, aggregation, or independent excerpts.
- Reconcile candidates against definitions, scope, dates, exceptions, amendments, cross-references, and precedence rules relevant to the task.
- Recover complete operative language from the original variable, then re-slice it to verify quotes, offsets, qualifiers, and conclusions.
- Call
SUBMIT(...) once with exactly the requested output fields. If the evidence is absent, ambiguous, or insufficient, state that instead of forcing the requested count. Do not SUBMIT an entire large llm_query or llm_query_batched blob; keep declared answers within the Turn output character budget.
Treat sub-model output as a candidate, never as source evidence. Respect the Turn's call and output budgets; reduce excerpts before making another call. Do not create chunk files, indexes, or paging state for ordinary sources; paging is a separate corrective path only if the whole-value benchmark fails.
Exact retrieval
For an exact quote, find the text with Python, retain its source offset, and compare the final excerpt byte-for-byte or character-for-character with the original input. If the evidence is absent or ambiguous, report that instead of inventing a quote or speaker.
1---2name: long-context3description: Use bounded retrieval to analyze large documents, transcripts, code, or datasets.4---56# Long-context analysis78Keep large inputs in variable space. Inputs may come from the user query, committed Session History, Skill resources, Attachments, a public URL, or Session Workspace. Explore them with bounded Python operations, send only relevant excerpts to sub-model calls, and verify every reported fact against the original variable before submitting.910## Analyze11121. Inspect variable names, types, lengths, and small previews. Do not print a whole large value. For a relevant URL, call `fetch_url` once and keep its returned `content` in a variable.132. Locate candidate regions with deterministic searches, indexes, regular expressions, or bounded slices.143. When the relevant regions are unknown, scan every bounded chunk for structured candidates before reducing them. Query ranking may prioritize reading order, but it must not exclude unseen evidence.154. Call `llm_query` or `llm_query_batched` only on self-contained excerpts that include the question and their source offsets or source identifiers. Use `rlm_query` only for the rare selected subproblem that needs a fresh iterative Python investigation; it is not the normal route for extraction, counting, parsing, aggregation, or independent excerpts.165. Reconcile candidates against definitions, scope, dates, exceptions, amendments, cross-references, and precedence rules relevant to the task.176. Recover complete operative language from the original variable, then re-slice it to verify quotes, offsets, qualifiers, and conclusions.187. Call `SUBMIT(...)` once with exactly the requested output fields. If the evidence is absent, ambiguous, or insufficient, state that instead of forcing the requested count. Do not `SUBMIT` an entire large `llm_query` or `llm_query_batched` blob; keep declared answers within the Turn output character budget.1920Treat sub-model output as a candidate, never as source evidence. Respect the Turn's call and output budgets; reduce excerpts before making another call. Do not create chunk files, indexes, or paging state for ordinary sources; paging is a separate corrective path only if the whole-value benchmark fails.2122## Exact retrieval2324For an exact quote, find the text with Python, retain its source offset, and compare the final excerpt byte-for-byte or character-for-character with the original input. If the evidence is absent or ambiguous, report that instead of inventing a quote or speaker.
Qredence