KCAE.Reference:1 - Source contributions and comparison boundaries
This is a practitioner synthesis with constructed applications. Primary research and implementation documentation supply mechanisms and bounded observations; they do not establish empirical effectiveness of a KCAE installation. The profiles above link the exact contributions used. Primary web material was inspected on 3 October 2026. A retrieval date identifies this reading; it is not an assurance that a later implementation is unchanged.
| Source line | Receiving contribution | Deliberate boundary and return |
|---|---|---|
| Bates, 1989, the berrypicking account, especially the evolving-search model | Questions evolve and useful information is gathered by several techniques; informs KCAE.SEARCH and MEMORY. | Historical conceptual lineage. Does not rank modern retrieval engines. Return when the interface makes the original episode or changing question unrecoverable. |
| Introduction to Information Retrieval, ranked evaluation | A mathematical basis for ranked retrieval measures and their dependence on query/relevance judgements. | Historical algorithmic reference; component relevance is not whole-work benefit. KCAE.EVAL adds that receiving comparison. |
| BM25, RRF, HNSW, ColBERT and HyDE, linked in Profiles:1 | Implementable lexical, fusion, approximate-vector, late-interaction and generated-query mechanisms. | Distinct operations with historical evidence. Select current realizations by use; source figures and parameter choices are not universal defaults. |
| Contextual Retrieval, GraphRAG, RAPTOR and LazyGraphRAG, linked in Profiles:1 | Different retained context, levels of abstraction and preparation/query cost arrangements. | Serious alternatives, not an ordinal maturity ladder. Preserve originals and inspect update and rare-distinction losses. |
| TypeSafe documentation and cookbooks, linked in Profiles:2 | Bounded typed screening, progressive candidate inspection and error-driven question development. | Provider mechanisms and demonstrations. Do not inherit synthetic accuracy, English performance, thresholds or explanatory-method exclusions. |
| Typed-decision audits, linked in Profiles:2.3 | Comparative-design and logical-coherence failure questions. | Early bounded evidence. Neither universal failure nor universal superiority follows. Reopen with changed models and better matched comparisons. |
| AgentRetrievalBench, especially natural no-gold analysis and limitations | A concrete test of how retrieval rejection behaves on naturally unsupported requests versus artificial wrong-corpus controls. | Code-retrieval evidence, not downstream task success or a support-corpus rate. Include natural no-use cases in KCAE.EVAL. |
| Provider, VibeVM and storage documentation in Profiles:3; CodeNib v2 | Available access forms, source identities, runtime loading and update guarantees at their stated scope. | Interface and bounded implementation evidence. It cannot establish that an installed reader used the source or that users benefited. Requalify the relevant contract after a version change. |
The best arrangement is unresolved until the receiving workload and comparison make the alternatives decidable. This publication adopts a source-returning, condition-aware construction because it supports those comparisons and local repairs. It rejects a universal requirement to begin with lexical-only retrieval, a universal preference for embedding/graph retrieval and a universal prohibition of dynamic loading. It also rejects treating retained bytes as sufficient access or a high retrieval score as justified use. These choices follow from the differing operations and failure cases developed in the methods, rather than from a claim that every technology has already been tested together.