WeKnora synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From sources to usable knowledge
WeKnora is a knowledge application with three connected modes: RAG answers, model-directed multi-step agent work and an editable generated wiki. This account covers its shipped application, including the callers that save and reload state, rather than treating its turn-local agent engine as the whole system. The evidence is SRC-1, the Tencent/WeKnora repository at 234aa948db29c10f5ed4b7939285c7b3711c5d4c, with an applicability cutoff of 2026-10-15. Inspected code supports implemented pathways; documentation and prompts support attributed intent. No target execution, observed trace or causal experiment is registered. External model, database, parser and sandbox service internals are excluded. Whole-system scope does not mean every branch was inspected (boundary, Source register).
A source first becomes stored material, then retrieval-eligible state. In the inspected file route, creation stores a source and knowledge row before asynchronous processing. Enqueue failure can leave a returned row marked failed. Successful indexing can enable retrieval while enrichment is still pending. Thus upload acceptance, retrieval availability and complete enrichment are different milestones, none of which certifies source truth (RT-RTE-knowledge-admission).
The retained source, current chunk body, historical revisions and search representations are also distinct. Parent chunks can provide context without their own vectors; indexed child/flat chunks include title and heading context. A later authorized edit can save a new body before reindexing fails, clear its original source locators and leave generated-question vectors indexed but marked stale. A current displayed body therefore does not establish that every derivative is current (MEM-OBJ-chunk, MEM-RTE-chunk-edit).
Wiki processing turns source contributions and existing pages into editable Markdown with revisions and reference handles. The reducer can receive cited passages, but a short Details summary is an alternative when citations are unavailable. Stored source/chunk references aid inspection; they do not prove sentence-level support. A machine-write gate can reject excessive loss of table-row identities and defer contributions for another batch. User edits, reverts and explicitly allowed shrinking have exemptions. This is a preservation check, not a factual oracle. Revert restores selected content fields, not necessarily historical provenance or placement (MEM-RTE-wiki-update, EPI-RTE-wiki-guard).
Withdrawal likewise has several horizons. Source deletion can remove references or sole-source wiki pages immediately, while mixed-page prose depends on queued model retraction, which can fail and leave the body unchanged. Structural lint can remove stale metadata without rewriting prose. Personal-memory deletion uses a separate bounded tombstone mechanism. Neither pathway establishes immediate global semantic erasure (MEM-RTE-wiki-update, EPI-RTE-wiki-maintenance, MEM-OBJ-memory-tombstone).
From a request to model-visible evidence and action
An ordinary authenticated agent request enters an owner-scoped session, persists messages and starts asynchronous work. The application resolves the agent's model, scopes and tools, reconstructs history, stages attachments, recalls personal memory and invokes the engine. In each ReAct round, the model proposes an answer or tool call; Go code controls eligibility, argument validation, context rewriting and stopping. Results feed subsequent rounds. SSE events expose intermediate work, and completion handling persists assistant content and AgentSteps. A final answer, partial stopped content, fallback or error can all terminate the route; completion is not successful task fulfillment (RT-RTE-agent-turn).
RAG uses a different progression: application code selects an event pipeline for retrieval, reranking and answer generation. Foreground agent knowledge search instead lets the model request evidence. That tool scopes, deduplicates, optionally reranks, selects and budgets results before rendering them. Wiki reads similarly deliver bounded page bodies or a compact index; omitted pages may be requested again. The consumer sees selected values, not all indexed material or automatically dereferenceable pointers to everything omitted (RT-RTE-rag-answer, MEM-RTE-knowledge-read, MEM-RTE-wiki-read).
Tool authorization is separate from answer warrant. Direct and catalog-resolved MCP calls reach shared execution and call-time policy checks. Approval is conditional on the configured checker and execution metadata; some checker failures can be configured fail-open. Shell/files and native browser capabilities have separate registration conditions, and host shell has an entitlement exception relative to remote execution. MCP approval therefore cannot stand for universal effect approval or sandbox isolation. Remote retry also does not establish exactly-once external effects (RT-RTE-mcp-call, RT-RTE-sandbox-execution).
The shipped prompt asks the agent to verify exact quotations, numbers and code against underlying passages. Source-reading tools make that behavior possible; neither the instruction nor schema/approval checks establishes a universal independent truth check. Ordinary QA has no reference-answer oracle. The evidence supplies no actual answer instance from which to establish entailment, correct citation or successful compliance (RT-CLM-evidence-verification, EPI-CLM-cited-answers, EPI-OBJ-chat-response).
What survives a turn and changes later work
WeKnora implements more than raw conversation storage. Selected user messages automatically feed personal-note extraction; resolved recurring topics can become interests; completed-answer references update document-familiarity counters. These outputs have later consumers. Notes enter automatic recall or requested memory search; interests and familiar titles supply query-rewriter background; familiarity counts multiply candidate scores in the inspected RAG rerank path. This is an implemented trace-to-durable-state-to-later-behavior pathway, not merely an API that might have a consumer. The unresolved question is whether those changes improve future action: counts record repeated selection, not satisfaction, and no benefit comparison is registered (MEM-RTE-personal-write, MEM-OBJ-memory-topic, MEM-RTE-retrieval-conditioning).
Personal admission is not uniformly human-reviewed. The extraction model proposes operations and supplies the inferred flag. Flagged non-explicit/non-manual inferences become pending; ordinary unflagged extractions and recurrence interests can become active without per-content confirmation. The prompt explicitly permits a non-entailed personal guess, but no actual such output is observed. Confirmation/rejection can change eligibility without recording the evidence or truth criterion behind the caller's decision. Source-message links preserve attribution, not the extractor's argument (EPI-RTE-memory-extraction, EPI-RTE-memory-write, EPI-RTE-memory-review, MEM-OBJ-memory-item).
Automatic recall selects a small resident block and query-ranked situational notes, with optional embedding selection. Requested memory search provides a separately targeted, larger read. Both wrap statements as background data, not permission to act; reliable model compliance remains unobserved. Timed or forced maintenance can expire, demote and model-merge notes, retaining supersession links but not merge reasoning. Starting chat or pressing consolidate is not approval of each admitted sentence (MEM-RTE-personal-recall, MEM-RTE-memory-search, MEM-RTE-memory-review).
Conversation continuity has a separate chain. Compaction uses an isolated model request and retains a summary plus recent material; failure can produce a raw archive instead. Eligible history checkpoints are saved on messages and actually reloaded by later turns. Checkpoint coverage controls which earlier turns are excluded, while filtered replay and budgets can omit further history. Tool-result previews are lossy and need not alter the raw ToolResult saved for history. These paths qualify as implemented trace-fed continuation, but neither checkpoint retention nor a smaller window proves semantic preservation or improved capacity (RT-RTE-context-rewrite, memory annotations).
Optional Q&A passage indexing supplies another search representation, containing the user question and assistant answer with thinking removed. The analyst records establish message-search read-back, not verified truth or automatic prompt injection. Report verification additionally identifies an optional conversation-search tool returning escaped, truncated prior Q&A while excluding the current session. That consumer must not be lost, but its omission from the partial inventory prevents exhaustive comparisons of conversation read-back and budgets (MEM-RTE-chat-index, MEM-OBJ-chat-passage, report verification limits).
Capability setup has a stronger delivery gate than an answer
The representative workflow is remote skill dependency installation, selected because a single coordinator connects delegation, rejection, bounded repair and durable delivery. It is not a representative quality test for all QA. The service seeds an accepted archive into a maintenance sandbox and briefs an installer engine. Model completion only permits verification: an error-free, complete engine state must precede tree, dependency, script and runtime-report checks. Explicit report blockers or missing declared commands can reject admission. Omitted requirements and meaningful end-to-end skill behavior remain beyond that partial oracle (RT-RTE-skill-install, workflow account).
Repairable structured findings can become another installer brief within a round budget. Sandbox effects and administrator guidance persist, but prior conversation history is not replayed. After verification, renewed ownership/lock and provider-account checks precede snapshot publication. Delivery occurs when the image pointer changes and later workspace resolution can select it, not when the model says installation succeeded. Pre-publication failure preserves the prior pointer; post-publication bookkeeping failure is not automatic rollback. The background install is not itself durable across restart. These are implementation-level admission and recovery distinctions, not an observed successful installation (workflow account, RT-BAP-skill-image).
Assessment and what would change it
For later personalized QA, the discriminating contribution is durable trace-fed context and retrieval conditioning, alongside both automatic recall and on-demand reads. For evidence-sensitive work, the corresponding constraint is that availability, reference handles, relevance scores and structural gates do not certify factual acceptance. For capability setup, the distinguishing contribution is a verifier-controlled publication boundary; its unresolved question is functional sufficiency and reliable later execution, not whether a pointer has an implemented consumer.
Learning as improved capacity for future action remains unestablished; the supported partial result is the retained update and later-consumer chain. Content-directed criticism of a consumed explanatory theory is not established by extraction, score adjustment or setup repair. The runtime does not supply a formulated-theory assessment sufficient to declare theory-builder conditions 1–4 met: localized theory content, content-dependent consumption, explicit criticism with blame placement, and retained criticism shaping iteration each remain unestablished as theory-specific properties. This is not a system-wide absence finding. The records also do not establish reflection's two causal directions through a self-representation or role-complete autonomous theory building. Computation owns selected maintenance and repair decisions; MCP approval and caller decisions remain conditional human-capable roles. Those operational facts neither establish nor refute every broader autonomy claim (RT-RTE-skill-install, epistemic bounded conclusion, report verification).
The installer implements evidence-responsive repair and an admitted change to later executable availability. Personalization similarly retains changes with consequential later consumers. These contributions are not evidence of exercised self-improvement: subsequent operation causally depending on the change is unobserved, and successful improvement has no outcome evidence. Installation alone supplies neither the complete attribution chain nor theory criticism (RT-RTE-skill-install, RT-BAP-skill-image, MEM-BAP-retrieval-feedback).
The dataset evaluator supplies reference-relative measurements, not that missing comparison. Its fixed RAG pipeline differs from open agent work. Content-to-passage mapping can drop unmatched retrievals before precision computation; lexical response metrics do not certify truth. Metrics reach a result caller without an established automatic successor-admission link (RT-RTE-dataset-evaluation, EPI-RTE-dataset-metrics).
Linked before/after traces showing the retained update, actual later selection and changed action would establish exercised pathways more strongly. Objective-specific controlled comparisons would be needed to attribute improved performance, with component variation for component effects. A complete functional skill test would strengthen setup acceptance; independent evidence-consuming factual acceptance would change the warrant assessment. Further source inspection could resolve named alternate routes but would not substitute for operational or benefit evidence. No additional frozen-source reads were made for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operation or causal comparison; external service internals excluded | SRC-1; boundary, RT-RTE-agent-turn, RT-BAP-skill-image, MEM-BAP-retrieval-feedback | Pinned shipped implementation and doctrine, not a deployed run | Successful delivery, provider compliance, deployed isolation/durability, exercised self-improvement and improved future capacity | Provenance-linked execution traces plus objective-specific intervention/comparison evidence; deployment/provider contracts for external guarantees |
| Additional optional conversation consumer is outside the analyst inventory's proven endpoint | SRC-1; MEM-RTE-chat-index, MEM-OBJ-chat-passage, report verification limits | Index/message-search chain; verifier inspected optional hybrid conversation-search tool | Exhaustive on-demand conversation-consumer and delivery-budget comparisons; automatic injection or demonstrated reliance | Accepted route-level records covering tool registration, owner/session selection, transformations, budgets and later model consumption; traces for exercised reliance |
| Some comparison units group distinct parts/consumers | SRC-1; RT-BAP-skill-image, MEM-RTE-retrieval-conditioning, profile verification limits | Grouped skill instructions/executables and familiarity rerank/rewriter findings within partial axes | Fine-grained unit counts and complete part-by-consumer comparisons, not the supported existence findings | Independently separated operative-part/consumer classifications with supporting records and coverage |
| Included alternate producers, payloads and consumers remain incompletely inspected | SRC-1; MEM-OBJ-chunk, MEM-RTE-wiki-update, RT-RTE-sandbox-execution, RT-RTE-mcp-call, memory coverage | Selected file/chunk/wiki/personal/history routes; connector/manual/FAQ, parser alignment, enrichment/graph/multimodal, wiki map/taxonomy/issues, attachments, workspace, approval/cache and client alternatives unresolved | Complete storage, lineage, authority, admission, evaluator, authorization or memory-route inventory; system-complete negatives | Same-pin producer-to-consumer records for each named alternative, including opaque payload boundaries, controls and maintenance |
| Semantic warrant and confirmation are narrower than operational admission | SRC-1; EPI-RTE-memory-write, EPI-RTE-memory-review, EPI-RTE-wiki-rewrite, RT-CLM-evidence-verification | Model proposals, conditional inferred/pending gate, caller confirmation, structural wiki checks and verification instructions | Universal human review, reliable inference classification, source fidelity, proposition-level acceptance or reason-preserving theory criticism | Candidate instances with source support, recorded evaluator evidence/criteria/intended use, and criticism/revision/consumer links; caller-decision evidence for human approval |
| Withdrawal, revision and retrieval freshness have separate failure horizons | SRC-1; MEM-OBJ-memory-tombstone, MEM-RTE-chunk-edit, MEM-RTE-wiki-update, EPI-RTE-wiki-maintenance, EPI-RTE-wiki-guard | Bounded tombstones/cache fallback, saved-body/reindex divergence, queued retraction, lint and selected-field revert | Universal non-reappearance, current derivatives, immediate global erasure or every-change historical rollback | Complete invalidation/transaction/recovery contracts across derivatives and historical stores, plus failure/retry traces |
| Context transformations can omit material and rationale | SRC-1; RT-RTE-context-rewrite, MEM-RTE-personal-recall, MEM-RTE-wiki-read, MEM-RTE-memory-review | Summary/raw fallback, filtered replay, previews and budgeted note/wiki delivery | Complete evidence preservation, full timeline audit, semantic equivalence or automatic recovery of every removed value | Linked producer/consumer payloads, coverage checks and explicit reload routes for omitted material; fidelity assessment against raw inputs |
| Model configuration is not an immutable parameter identity | SRC-1; RT-CMP-chat-model, RT-CMP-embedding-model, RT-CMP-rerank-model, RT-CMP-vlm-model | Inference adapters, names/URLs/dimensions; provider mutability uninspected | Exact weight provenance, stable model behavior or absence of provider-internal changes | Immutable provider/model identities and versioned deployment evidence; relevant provider update contracts |
| Installation gate and recovery are bounded to inspected remote setup | SRC-1; RT-RTE-skill-install, workflow account | Declared prerequisites, bounded repair and pointer publication; host equivalence, reaper, explicit rollback and provider snapshot semantics unresolved | Complete functional correctness, crash-safe exactly-once publication, equivalent host guarantees or arbitrary-effect rollback | End-to-end functional criteria, complete host/reaper/rollback records and provider contracts; restart/concurrency traces |
| Evaluation is mapping- and reference-relative | SRC-1; EPI-RTE-dataset-metrics, RT-RTE-dataset-evaluation | Mapped passage IDs, selected lexical metrics and fixed evaluation RAG; full formulas/reference fidelity and durable retention unestablished | Precision over all returned chunks, general answer truth, open-agent quality, causal benefit or automatic evaluator-driven improvement | Accounting for unmatched results, complete metric/reference validation, observed comparisons and any consequential successor consumer |