qredence/fleet-rlm/long-context-search

Long-context search

Use this when a selected dataset is too large to print or the answer depends on scattered passages in that file, such as a log, transcript, or long document. For other workspace files, inspect workspace.tree() and read the relevant files in bounded pieces; the context methods apply only when a dataset was selected.

Steps

1. Measure first: workspace.stat(path) gives the size; workspace.tree() lists the files.
2. A selected dataset arrives as context. Use context.preview(),

context.lines(), context.chunks(chunk_bytes) and context.search(query, limit). search returns (line_number, line) pairs.

1. Narrow with code before using models. Use exact keyword or regex filters,

and do counts and group-bys in Python. Print the counts.

1. For semantic judgments over many passages, batch them:

``python prompts = [f"Does this line report a failed payment? Answer yes or no.\n{line}" for _, line in hits] answers = llm_query_batched(prompts) ``

For a fixed set of labels per item, jev_decide_batched(items, question, choices) returns one typed decision per item.

1. Keep evidence: store (line_number, excerpt) for every claim and cite the

line numbers in the answer.

1. Count and aggregate in code from the decisions, print the totals, then

SUBMIT from those variables.

Pitfalls

• Do not print a whole large file; it is truncated and wastes iterations.
• lines() raises on lines over 1 MB; use chunks() for binary or minified files.
• Sub-LLM calls share the Turn budget; filter first, then ask.