KCAE.SEARCH - Continue a Search as the Question Develops

Type: Method pattern Status: Stable

KCAE.SEARCH:1 - Problem frame

Use this when a question has no sufficient known source, when a first result leaves a consequential gap, when new reading changes what should be sought, or when the answer concerns a collection rather than one passage. The governed object is one bounded inquiry through available access routes. The result is useful inspectable material, a supported synthesis within a declared corpus scope, or a precise unresolved/access limit, with enough history to continue without repeating a failed interpretation. An already sufficient source or result is a direct exit.

KCAE.SEARCH:2 - Problem

A fixed-query loop can repeatedly find the same plausible material. A free-ranging agent can spend its budget without reducing the important uncertainty. Both can return “nothing found” while having considered only one vocabulary or one subset. The next search must follow the receiving question and what the previous reading actually changed. A broad answer additionally needs a way to combine source claims under a declared coverage basis; finding several relevant passages does not perform that aggregation.

KCAE.SEARCH:3 - Forces

Preserving the original question protects its meaning, while inquiry legitimately changes it. Broader exploration improves opportunity and consumes scarce reading resources. Early useful candidates justify focused reading, but can anchor interpretation. A useful stopping decision needs the limits of the available routes, not a claim of universal absence.

KCAE.SEARCH:4 - Solution

KCAE.SEARCH:4.1 - Keep the question, facts and variants separate

Retain the user’s wording or source episode, the receiving result, known facts and current unknowns. Add a corpus-language translation, a description of the missing result, and a rival interpretation when those can expose different material. For CedarBench, “buy faster workers” remains the request; “prevent duplicate resend” is a search hypothesis supported by the morning correction episode, not an established cause.

Generate variants from relations as well as nouns. Ask what operation might produce the missing result, what condition might defeat it and which result an already found method requires. A hypothetical answer can provide retrieval vocabulary, but any invented product, causal link or configuration remains outside the case facts. When an actual reading changes the question, state the change and preserve the unresolved part of the earlier one.

KCAE.SEARCH:4.2 - Select the next route by the missing contribution

Choose a route whose selection ground can plausibly reach what is missing. Known identifiers favour exact lookup. Vocabulary mismatch favours corpus-language variants or semantic retrieval. A missing definition favours structural expansion. A global synthesis can need several regions or multiscale summaries. An index blind spot or a newly changed region can justify original-block inspection. The available arrangement may supply only some of these routes; lack of a route is a capability limit.

Keep a small search state: original/current question, inspected source units and editions, useful contributions, disputed assumptions, pending source returns, covered extent and remaining resources. Avoid retaining every verbose search result in the principal reader’s context. The state guides the next action; an authorized detailed log can remain outside that context for diagnosis.

For a question with several needed aspects, state what each contribution must establish and mark which remains missing or disputed. More documents about an already answered aspect do not supply another one. Compare new material by the contribution it adds, not only by whether its identifier is new. For example, ten resend paragraphs cannot replace the missing destination configuration; once the governing rule and configuration are sufficient, another synonym search needs a separate reason to continue.

The next action should have an explicit expected contribution: “Open the compatibility note to determine whether protocol v1 supports deduplication,” rather than “search more.” When the required fact is the customer’s installed version, use a permitted configuration source or ask the customer. No amount of searching the public manual will establish that local fact. This is a return from finding literature to obtaining case data.

KCAE.SEARCH:4.3 - Read promising material and test the interpretation

Open the source around decisive hits through KCAE.SOURCE. Expand through the condition, exception, definition or example that can alter the proposed contribution. Preserve conflicts instead of choosing the more convenient passage. A preliminary score can prioritize this reading; KCAE.ASSESS turns it into a supported or unresolved contribution.

Actively seek a discriminating countercase when two readings would lead to materially different actions. If a retry paragraph appears to allow unattended resend, inspect restrictions on receiver versions and receipt state. That is targeted challenge, not a demand to read every related publication. If the interpretation survives and the receiving result is available, stop. If it fails, use the discovered condition to formulate the next question.

KCAE.SEARCH:4.4 - Escape a consequential blind spot

When the query family is unfamiliar, the user rejects the interpretation or the necessary result remains missing, inspect the selection assumptions. Was the corpus reduced to one department? Were examples excluded? Did every query use the same translated diagnosis? Change the relevant assumption and choose a route not bound by it. The alternative need not be expensive: a glossary term, source table of contents or colleague’s exact reference may suffice.

For a direct semantic pass, use the independent enumerator from KCAE.INDEX:4.5. Partition by actual accessible source units, not by the failed relevance category. Allocate a bounded initial portion; retain source order or sampled strata, examined extent and reasons for further expansion. Inspect promising units fully and search for missing links across them. A new question can invalidate a prior “not useful” screen, so cache that result with its criterion and question rather than suppressing the unit forever.

If only part of a corpus was examined, return that extent and the practical consequence. “No adequate current automation rule was found in the inspected public manual and release notes; the supplier bulletin was inaccessible” is actionable. “There is no rule” is a stronger conclusion unsupported by that search. A complete deterministic search can establish absence of an exact string from an exact corpus; it cannot establish absence of every relevant meaning.

KCAE.SEARCH:4.5 - Spend and stop at the receiving result

Allocate resources to finding, assessment and the reading still required for correct use. Reserve enough for the final necessary source context; spending all capacity on candidates prevents application. Use actual tool/model limits, including latency and rate limits. Stop or narrow the inquiry when a required channel is unavailable, a hard budget is reached or additional accessible work is unlikely to change the next decision enough to justify its burden.

The judgement can be qualitative. Compare the next affordable action with continuing from the present result: what uncertainty could it remove, what action would change and what would the inspection cost? If a numerical value-of-information calculation is used, obtain its probabilities and costs independently; a retrieval score cannot stand in for those values. KCAE.USE supplies the decision context, while C.11.DUA supplies the general burden question.

Return the source candidates and why they matter, the inspected conditions, unresolved premises and any coverage limit that changes use. A short reply can preserve all of these where the case is simple. Distinguish “sufficient material found,” “current candidates rejected,” “missing fact,” “access constrained” and “budget stopped.” They imply different continuations and should remain different in a machine interface.

KCAE.SEARCH:4.6 - Construct a bounded corpus-wide synthesis

Use this branch when the requested contribution concerns a collection: its themes, objections, changes, contrasts or distribution of stated positions. First distinguish three results. An exemplar shows that a particular contribution occurs. A thematic overview organizes identified contributions and their differences over declared material. A frequency claim counts a defined feature in a defined unit population. Finding a good example answers the first question; repeatedly retrieving it cannot answer the other two.

Set the population and the claim before choosing the aggregation route. Specify membership, time or edition rule, permissions, and the unit about which the answer will speak. Eight current submissions, twelve stored files and eight submitting organizations can be three different populations. Decide whether earlier editions are historical evidence to compare or superseded copies to exclude from the current count. Distinguish documents that repeat one underlying report from independently produced evidence. For themes, state what “main” means in the receiving use: commonly recorded, explanatory of a contrast, or consequential to the decision. Frequency alone need not determine importance.

Obtain an inventory from the source system or KCAE.INDEX’s enumerator. Keep each included unit’s source identity and selected edition, its assigned reading portion, and whether the relevant content was read, excluded by a stated rule, inaccessible or still unexamined. If only a provider’s selected hits are available, that is the available set; do not call it the complete collection. A broad question can legitimately produce an exploratory overview of selected material, provided the answer keeps that scope.

Choose a route that can cover the required material at an affordable cost. For a small dossier, a qualified reader can read every included unit, record its question-relevant claims with source returns, and compare them directly. This avoids index and summary preparation and is a serious choice for infrequent inquiries. If the material exceeds one reader’s working capacity, partition the enumerated units into bounded reading portions. Give each portion the same question and inclusion rule. Split a long document without losing its unit identity; preserve the necessary cross-boundary context and reconnect claims that span portions. This direct partitioned map/reduce route is available without an entity graph or precomputed summaries.

For a large, repeatedly queried collection, use prepared summaries or communities where their saved query effort earns their preparation and update burden. Choose the regions or hierarchy levels to inspect and resolve their membership back to source units. A parent and its child summaries can describe the same evidence; several graph communities can reach one underlying document. Selection across levels, including RAPTOR’s collapsed-tree profile, can find useful material without inspecting the entire population. A query-guided route such as LazyGraphRAG can allocate reading to promising communities, but the retained selection boundary still limits the answer. KCAE.Profiles:1.3 compares these constructions. Use independent original inspection to investigate consequential regions or distinctions the prepared view may omit.

Map source material to claims that can be combined. For each reading portion, obtain the proposed answer-bearing claim, its source units and exact passages, the relevant entity/time/condition, and any contradiction or unresolved interpretation. Preserve which words are the source’s position and which relation the reader inferred. One portion may supply several themes, a counterexample, an exception or no relevant claim; do not require one winner or one positive answer. A “no claim” screen remains revisable when the question or coding criterion changes.

Carry lineage through intermediate summaries: a claim refers to its contributing source-unit set, not just to the name of the summary that repeated it. When an operation cannot preserve that lineage, use its summary to locate original passages before relying on the aggregate. Keep unread or excluded portions visible alongside positive results. A polished partial answer with no account of what it left out is insufficient input for a claimed whole-corpus conclusion.

Reduce by meaning and source support, not by the number of summary votes. Align claims that concern the same object, time, condition and predicate. Merge genuine paraphrases while taking the union of their underlying source-unit sets. Do not merge contrary positions or a conditional exception into the majority wording. Group the resulting claims into themes that answer the receiving question, then explain both recurring relationships and consequential differences. If an initial theme does not explain a substantial contrast, refine it and reopen the source portions whose interpretation can change; merely relabelling the final paragraph leaves the earlier coding unchanged.

For a descriptive count, define the predicate and count each eligible population unit once for that predicate. A unit may belong to several themes; disclose that those categories overlap rather than forcing their percentages to total one hundred. Retain unknown or unread units in the account of the denominator. Do not divide a relevance-selected hit count by the whole collection and call it prevalence. A claim about a wider population needs a justified sampling/estimation method and its assumptions; otherwise return the observed corpus count or a qualitative overview. The number of documents repeating a claim also does not by itself establish its truth.

Protect consequential minority evidence before compressing the answer. Maintain the supported objections, contrary cases, important qualifications and unresolved conflicts that could change the receiving decision, even when they are rare or rank poorly. Ask the relevant domain reader which differences have that consequence when ordinary interpretation cannot settle it. Reserve enough final reading and answer space for those differences, or explicitly narrow the promised answer. If intermediate claims exceed the reduction window, reduce them in further bounded groups while preserving source sets and these outstanding differences; keep the originals obtainable for the final comparison. This controls the next reading operation, not semantic loss by fiat.

Return an answer whose scope survives the final wording. State the population and edition basis, the supported themes or counts, material counterevidence and the unexamined or inaccessible remainder. Link decisive generalizations through their intermediate claims to originals. Verify that words such as “all,” “most,” “typical,” “increasing” and “consensus” have the required comparison or counting basis. A minority warning can be important without being typical; silence on a question is not agreement with another submission.

Stop when the intended bounded answer has adequate inspected support, consequential conflicts have dispositions, and the remaining affordable inquiry would not change that receiving result enough to justify its cost. A thematic overview can stop with a named uncovered region when the recipient accepts that narrower use. A complete descriptive count must instead obtain the necessary unit judgements or report the unresolved count; a representative estimate needs its own sampling basis. When the hard budget ends first, return a partial synthesis and the particular next region or ambiguity that matters. Reading every unit establishes an inspection extent, not guaranteed recognition of every possible meaning. KCAE.ASSESS checks the proposed answer against this basis before KCAE.DELIVER carries it to the recipient.

KCAE.SEARCH:5 - Archetypal Grounding

KCAE.SEARCH:5.1 - An evolving local question

The first CedarBench query finds worker-capacity instructions. The retained episode includes duplicate removal, so the second query concerns receipt reconciliation. The source then distinguishes destination protocols. The next useful operation is a permitted local configuration read. It returns v1, redirecting the answer from automatic resend to reconciliation. If the configuration channel is unavailable, the result is a conditional answer and one precise information request. The search has made useful progress without pretending that it established the missing fact.

KCAE.SEARCH:5.2 - A dossier’s objections, recurring themes and minority condition

Consider a constructed consultation on a research archive’s proposed deposit requirements. The receiving question is: “What objections do the current submissions raise, and which conditions should the archive examine before adopting the proposal?” There are eight submitting groups, twelve source files because four are superseded editions, and two overlapping generated summaries. The declared population is the eight groups’ latest submissions at the closing date. Earlier editions remain available for historical comparison; the summaries are access views. Neither adds another current respondent.

The eight short submissions fit direct reading in four portions of two, with the same question and source-role instructions. The reader opens every selected edition and returns the following claim map. These invented labels abbreviate source-addressed readings; an installed system retains the actual addresses and qualifiers.

For the descriptive count, a direct effort objection means an explicit rejection of the proposed metadata workload. A conditional acceptance such as R8 is coded separately and its unresolved condition remains in the overview. For this coding, the direct metadata-objection set is {R1, R2, R3, R4, R5}; storage cost is {R1, R4, R6}. A summary of R1–R4 and another of R3–R8 both mention metadata. They supply overlapping paths to those sets, not two more respondents or two independent confirmations. An older R5 file likewise cannot increase the count. The union operation preserves five current direct metadata objections and three storage objections out of eight, with R1 and R4 in both categories.

The resulting bounded answer is: direct objection to metadata effort is the most frequently recorded objection category in these eight current responses (five of eight), followed by storage cost (three of eight). Those counts describe the submissions, not all potential contributors or the strength of an objection. R8 identifies a template condition under which the effort concern may be resolved; R7 identifies a disclosure problem that an effort/cost majority does not answer. The archive needs to inspect those conditions before treating the aggregate as support for one uniform policy. Each theme returns to its listed sources, and the important contrasting statements return specifically to R7 and R8. No inference of consensus is made from the other groups’ silence about locations.

Now change one condition. R5 submits an authorized replacement before the closing date: after trying the supplied template, it withdraws the effort objection. Reopen R5’s coded contribution and the aggregate that used it, leaving the other inspected claims intact. The current metadata-objection set becomes {R1, R2, R3, R4}: four of eight, so “more than half object” is no longer justified. The template contrast strengthens in a stated way, but the disclosure concern remains. The old five-of-eight result still describes the earlier snapshot; keeping the old file or an unrefreshed summary cannot make it current. If the replacement’s meaning cannot be read, report four confirmed objections and R5 unresolved, rather than silently counting the previous position.

For this small, infrequently queried dossier, direct reading is the selected design: it supplies every current unit without a maintained graph. The four-portion map/reduce construction is a capacity variation of the same source route. A large archive with many repeated cross-document questions can justify prepared community summaries or query-time graph exploration, after comparing their full cost and ability to preserve the minority and changed-edition cases. Their value is not established by needing fewer summaries than original files. The stopping basis here is the inspected eight-unit population, resolved edition membership, traceable coding and stated conditions; no claim of universal thematic completeness is required.

KCAE.SEARCH:6 - Bias-Annotation

An agent can overfit its first diagnosis and mistake additional supporting hits for independent confirmation. Preserve rival formulations and the user’s rejected interpretation. Search logs also omit unseen opportunities; KCAE.EVAL needs separately authored cases to expose them.

KCAE.SEARCH:7 - Conformance Checklist

Is the original concern recoverable? Does each expansion seek a needed contribution or test a consequential interpretation? Can the search leave the failed selector? Are case facts kept separate from generated hypotheses? Does the stop state accurately describe source coverage, inaccessible material and remaining uncertainty? For a broad answer, can the reader recover its population, aggregation units, source lineage and consequential counterevidence, and distinguish a theme from a frequency claim?

KCAE.SEARCH:8 - Common Anti-Patterns and How to Avoid Them

Repeating paraphrases of one assumed diagnosis preserves its blind spot; reformulate the missing result or inspect the original episode. Treating no hit as absence exceeds the route’s evidence; return its limit. Searching public documents for a private configuration fact wastes budget; obtain the fact from its proper source.

KCAE.SEARCH:9 - Consequences

Search becomes a sequence of intelligible contributions to work. It may end earlier with a sufficient result or continue farther when a different route has real value. The method preserves uncertainty and cannot guarantee that the next unexamined source would add nothing.

KCAE.SEARCH:10 - Architectural Rationale

The current gap connects earlier reading to the next operation. Preserving both the original concern and the evolving question avoids the extremes of a fixed query and aimless exploration. A truthful stopping limit supports continuation by another reader without reconstructing all search history.

KCAE.SEARCH:11 - SoTA-Echoing

Bates’s developed historical comparator already permits changing questions, several techniques and movement among sources. This pattern adopts that inquiry, rather than claiming iteration as its invention. Its engineering addition connects each unresolved contribution to available versioned routes, reading capacity, source-return operations and truthful stop states. A fixed retrieval call remains a smaller alternative when the need is stable and it suffices.

The contemporary BRIGHT-Pro study, section 6.3, provides bounded agentic-search failure evidence: rewritten queries can repeat unhelpful evidence, new documents can cover only one aspect, and exploration can continue after sufficient evidence was found. KCAE.SEARCH:4.2–4.5 addresses those distinct failures through contribution tracking, selector revision and a receiving-result stop. These are proposed repairs to compare in the intended workload, not a claim that this publication was tested in that study. Reopen the route policy when it repeatedly misses valuable material or spends effort without changing the result.

KCAE.SEARCH:12 - Relations

F.1 supplies question-relative source selection and an optional search policy in the cited edition. This domain pattern realizes an inquiry through the available retrieval arrangement. KCAE.INDEX supplies routes; KCAE.ASSESS judges inspected contributions; KCAE.COMPOSE generates missing-result queries; KCAE.DELIVER transfers the necessary result.

KCAE.SEARCH:End