search-strategy

Operates under retriever:routing. If the user's intent actually fits a different tier — another retriever:* skill, a Tier 2 slash, a Tier 3 tools.py subcommand, or (last resort) direct DB access — stop and re-route against the ladder before continuing.

Retriever Search Strategy

Read search-strategy.md before presenting search results or constructing filters.

Rules

• Prefer SQL-like --filter "<expression>" filters over tuple-style field/operator/value input.
• The filter grammar applies to Retriever's logical document fields, including supported virtual fields such as production_name, is_attachment, and has_attachments.
• Default to relevance sorting for keyword queries, Bates ordering for Bates lookups, and date_created desc with nulls last for browse/filter-only views.
• Use Retriever's canonical search CLI flags: --filter, --sort, --order, --page, --per-page, --columns, and --mode.
• Map "show N" style requests to --page 1 --per-page N; do not invent --limit.
• Use canonical field names such as date_created; do not invent variants like created_date.
• For persistent browsing, use the slash surface: /search, /bates, /filter, /dataset, /from-run, /scope, /sort, /page, /next, /previous, /page-size, and /columns.
• Bare slash commands are read-only state inspection when supported: /scope, /dataset, /sort, /page, /page-size, and /columns.
• Use list subcommands for discoverability: /scope list, /dataset list, /sort list, and /columns list.
• Unless the user explicitly asks for a different layout, show document results using the active display column set. The default when no override is present is content_type, title, author, date_created, and control_number.
• Treat "show", "show me", "list", "display", "browse", "which documents", and similar listing verbs as requests for the standard /search table output, not as permission to improvise a custom list format.
• When the user asked to show/list/display/browse documents or to see the table itself, prefer the slash/session browse surface so the saved /page-size, /columns, and /sort state apply, then forward the tool's rendered_markdown as the entire reply. Do not add any text before or after the table.
• Do not exceed the active /page-size when showing document rows unless the user explicitly asks for more or changes the page size.
• In compose mode, keep summaries short by default. Use one short paragraph unless the user asked for a detailed write-up or the material genuinely needs a second paragraph for chronology or contrast.
• In compose mode, anchor the main answer to the top matching documents instead of broad collection-wide storytelling. Name the most relevant control_number items when they support the conclusion.
• Do not add side remarks about unrelated hits, dataset noise, or collection quality unless the user asked for that analysis or it materially changes the answer.
• When the user asks to show files, documents, or attachment children, make every shown result clickable with its preview/open link and keep the standard /search schema unless the user asked for another layout.
• The Title cell is always the clickable cell in the standard format; do not create a separate Link column.
• Use the tool-returned preview path when one exists, and fall back to the source/native path when no generated preview exists.
• When control_number is present in the active column set, keep it in its own column rather than embedding it in the title link.
• Show the active scope, sort, and page clearly so the user knows what they are looking at.
• Use Retriever's default compact search payload first; rerun with --verbose only when you need attachment child rows, alternate preview targets, or extended metadata that compact mode omits.
• When a search or DB step yields multiple independent follow-up paths, issue those reads/searches in one assistant turn instead of one-by-one.