Skip to content

Knowledge Platform — Workflows

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Source onboarding and snapshots

A Workspace admin registers a source (kind, reference, extractor, classification, ACL mapping, sync policy) or the project's source scope (BC-01) creates it. The scheduler takes snapshots per policy or on change notification; a snapshot is immutable and flows Taken → Parsed → Chunked → Indexed → Published, superseding the previous one only after ACL propagation. A failed stage leaves the previous snapshot published and surfaces as Stale or Failing.

Ingestion to retrieval

sequenceDiagram
    participant I as KnowledgeIngestionService
    participant R as VectorMemoryService
    participant MG as ModelRouterService
    participant RI as RepositoryIntelligenceService
    I-)R: ChunksPublished (embedding profile version)
    R->>MG: Embed through the embedding profile
    R->>R: write vectors and text items under the scope filter, graph extraction
    R-)I: IndexRevisionReady
    I-)RI: SourceSnapshotTaken (repositories)
    RI->>RI: build symbols, dependencies, calls, ownership, tests
    RI-)R: RepositoryIndexed
Hold "Alt" / "Option" to enable pan & zoom

Index revisions are pinned to an embedding profile version; a profile change or a restore rebuilds the next revision from published snapshots while the old one serves.

Context assembly per attempt

The runtime calls Assemble with the attempt's scope and budget; the planner allocates sections, retrieval and graph expansion return cited candidates, promoted memory and repository state are added, ranking, deduplication, compression and budgeting produce the assembled context, and the snapshot records every included and dropped item. A degraded index marks the snapshot Degraded; mandatory sections that do not fit make it Failed rather than truncated.

Artifact chains

Every stored content is an artifact in a version chain with lineage to its inputs; RunCompleted seals a run's artifacts; validators attach verdicts; previews are projections; retention and export follow the project's data lifecycle.

Memory learning loop (original handoff §32, §69)

flowchart LR
    RC["Run completed / attempt outcome"] --> O["Observations"]
    O --> C["Candidate memory"]
    C --> K["Classification<br/>temporary · episodic · semantic candidate · procedural candidate · discard"]
    K --> E["Evidence check and trust tier"]
    E --> P{"Policy"}
    P -- "episodic, run or project scope" --> Pr["Promoted"]
    P -- "semantic or shared" --> Rv["Review (approval of class publication)"]
    Rv --> Pr
    Rv --> Ar["Archived"]
    Pr --> Re["Reinforce on use / decay when unused"]
    Re -- "procedural, reinforced" --> Sk["Skill candidate proposal (BC-04)"]
Hold "Alt" / "Option" to enable pan & zoom

Deletion lineage (D31, XL-14)

Deleting a source or closing a project produces lineage records for every derived chunk, vector, graph fact, context snapshot citation and memory evidence; the propagator executes them within 30 days, exports first when the tenant chose export, and writes deletion evidence to audit.