HippoRAG reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

Boundary and evidence

The four reports share the whole-system boundary and its Source register. Their Python, MCP and plugin accounts concern different responsibilities within that boundary, not different targets. Implementation, doctrine, operation and benefit remain separate. No supplied report establishes an observed target execution or causal benefit. This reconciliation uses the supplied members only; no source reads were needed. No correction answers or declined objections were supplied.

Duplicate identity and retained parts

Amendment: EPI-OBJ-passages is superseded by MEM-OBJ-passage-text. Both identify the retained natural-language chunk-store content, supplied as caller passages and later returned unchanged from stored rows. Their shared get_row(key)["content"] passage establishes the same object, not merely similar roles. The epistemic inventory, acquisition/retention/selection ledger entries and bounded conclusion therefore concern the memory record's retained passage object. The epistemic finding that passage provenance does not verify source truth remains unchanged; neither the parent indexing route nor the per-call query container is superseded.

No other object identity requires supersession:

  • EPI-OBJ-openie-triples identifies relation content in extraction results and persisted rows. MEM-OBJ-openie-state identifies the retained state container, including passages, entities, provenance and loader behavior. Relation content is not identical to that whole container. The latter's omission of raw response metadata remains distinct from MEM-OBJ-response-cache, which retains raw responses for matching inference requests.
  • EPI-OBJ-normalized-facts describes normalized relation content across its stored representations and source mappings. MEM-OBJ-fact-text, MEM-OBJ-fact-embeddings and MEM-OBJ-graph-state distinguish text, vectors and graph consumers. The epistemic record explicitly distinguishes those representations; its findings do not become wrong when the memory parts are separated. It remains a valid content-level account, not an identity with any one part or a split supersession. MEM-OBJ-entity-text and MEM-OBJ-entity-embeddings likewise identify different operative material from relation tuples and their fact vectors.
  • EPI-OBJ-answer-thoughts is content within RT-OBJ-query-solution, not another declaration of the whole result container. The container also carries selected passages, scores and metadata. Neither is the retained passage corpus or the response-cache object.
  • MEM-OBJ-passage-embeddings records stored encoder output, not RT-CMP-embedding-model. MEM-OBJ-chunk-metadata records caller provenance/access metadata, not the compatibility fields of RT-OBJ-index-manifest. The remaining runtime components distinguish inference roles and offline engines; shared weights do not make those roles duplicate records.

Route connections without identity

EPI-RTE-openie-generation, EPI-RTE-triple-check, EPI-RTE-triple-admission and EPI-RTE-text-normalization separate generation, structural checking, disposition and normalization within RT-RTE-index-state. None replaces the complete index/delete/reload lifecycle. Their separate targets preserve the distinction between admitting processable structures and accepting relation truth. RT-RTE-offline-extraction supplies alternate staging and parser behavior; it prevents extending online rejection findings to all extraction modes.

EPI-RTE-fact-check separates model relevance judgment from deterministic remapping in EPI-RTE-fact-selection. Both are narrower than RT-RTE-retrieve-answer, which also covers ranking, passage delivery and QA. Candidate membership, relevance and truth remain different conclusions. RT-RTE-standard-rag is a separate dense baseline, not an alias for graph retrieval. RT-RTE-ircot-retrieval adds thought-directed current-call retrieval and termination, not durable thought admission.

MEM-RTE-response-reuse makes the runtime's brief cache finding explicit at wrapper scope across extraction, filtering and QA. It is neither identical to the complete retrieval route nor confined to it. EPI-RTE-metric-evaluation similarly spans evaluated modes rather than replacing one retrieval route. It supplies score production and reporting, not an acceptance or method-revision consumer.

WF-RTE-folder-indexing spans host preparation, confirmation, optional refresh, multiple service calls and final reporting. It is not identical to RT-BAP-plugin-skills, the instruction channel; RT-RTE-mcp-tools, the individual queued service operations; or the library lifecycle. EPI-RTE-host-check is a separate instructed answer-checking subpath of plugin guidance, not folder ingestion or a server-side semantic gate. These overlaps require no route supersession.

The epistemic claim records also remain distinct: EPI-CLM-memory-connections records public purpose, EPI-CLM-grounded-answer records the QA grounding claim, and EPI-CLM-call-guidance isolates cost/call guidance. Their implementing routes and the plugin authority path are not duplicate claims. RT-CLM-state-compatibility concerns checked configuration and identity, not any of those semantic or cost claims.

Convergence and integration-issue dispositions

Call guidance. The memory report's integration issues, the workflow's cost qualification and the epistemic claim comparison converge with the runtime's existing qualification. EPI-CLM-call-guidance, EPI-RTE-openie-generation and RT-RTE-retrieve-answer preserve the difference between skill statements and separate online extraction stages/model-backed filtering. This is a source doctrine/implementation mismatch reported consistently, not a disagreement between analysts. Cache hits and bypasses prevent substituting one universal actual cost figure.

Deduplication and provenance. MEM-OBJ-chunk-metadata qualifies the server's no-op description: identical text can replace metadata without adding embeddings or extraction. This does not contradict RT-RTE-index-state, whose missing-hash selection concerns new content work. The workflow's WF-RTE-folder-indexing preserves that narrower content-skipping claim. Source-ID refresh targets current metadata; it does not establish separate ownership of identical text from two sources. The memory issue therefore adds a limit rather than requiring a runtime supersession or correction.

Withdrawal and recovery. MEM-RTE-response-reuse and MEM-OBJ-response-cache establish why traced index deletion is not secure forgetting. The runtime claims deletion of retained index contributions, not coupled cache erasure. WF-RTE-folder-indexing adds the consequence of successful deletion followed by failed replacement. All three accounts distinguish local rejection/serialization from multi-file or whole-workflow atomicity; RT-CLM-state-compatibility supplies no rollback guarantee.

Meaning and authority. EPI-RTE-text-normalization supplies a collision deduction consistent with the memory annotation's warning that normalization can collapse surface distinctions. Original passages remain available; neither report claims all distinctions disappear from every consumer. Structural admission, retained graph support, query relevance and metric agreement are not interchangeable warrants. EPI-RTE-triple-admission, EPI-RTE-fact-selection and EPI-RTE-metric-evaluation keep their separate criteria and force. The bounded lack of a semantic acceptance route is not a repository-wide or provider-internal absence claim.

Host and service responsibility. RT-BAP-plugin-skills, EPI-RTE-host-check and WF-RTE-folder-indexing consistently retain host-interpreted guidance as claimed, service wiring as implemented, and actual host adherence as uninspected. Read-only registration and one service queue do not enforce content approval, host reporting or cross-process exclusivity. The workflow's returned count evidence does not establish semantic verification or a complete file-level audit.

Remaining uncertainty. The memory integration questions about alternate caches, backend persistence/concurrency, fixture provenance, weight identity and host behavior remain local limits. The other reports do not resolve them by inference from default-path wiring. Available but uninspected material remains distinct from inaccessible external internals, and neither establishes absent behavior. No supplied dependent finding requires a cited record to say something it no longer says.

No unresolved cross-report disagreement or required missing part was identified in these pinned members. The supersession above connects one duplicate without replacing either analyst's findings; all other overlaps are distinct parts, containers, claims, channels or route scopes. Evidence limits remain with their declaring records rather than becoming a single system-wide status.