Fabric handbook · Executor + Studio

RAG, Graphify and agent knowledge

Knowledge makes an agent better informed. It does not provide missing credentials, executable capabilities, certification or approval.

◇

Reference

Understand the different kinds of context

SystemWhat it contributesWhat it does not do
RAGRetrieves relevant excerpts from indexed, accessible sources with provenance.Copy entire repositories into every prompt or make old content current.
Conversation memoryRecent turns and compacted context for a workspace/session.Certify facts, approve effects or replace run records.
Agent observationsRetained evidence about actual stage attempts and outcomes.Turn a failed or interrupted stage into success.
SkillsPlaybook instructions and, where valid, typed capability contracts.Execute arbitrary downloaded prose automatically.
MCPA protocol for discovering and invoking provider tools under permission.Act as the RAG corpus or grant authority just by registering an endpoint.
GraphifyRevision-bound structural code relationships such as symbols and dependencies.Parse every business domain or prove runtime correctness.
Production GraphA read projection connecting code, assets, tickets, runs and evidence.Execute, approve, retry or acknowledge work.
◇

Reference

Add and configure RAG sources

  1. Open Settings → Knowledge. Inspect the current index status, last build time and source list.
  2. Under Custom evidence sources, choose Local path for an accessible file/folder or Web URL for a page/Google Sheet URL. Enter the exact path or URL and an optional useful label.
  3. Click Add source. Assign a collection that matches its domain; use Common knowledge for genuinely shared material.
  4. Configure the relevant connector first for protected Confluence pages or Google Sheets. Your browser being signed in does not authenticate the backend.
  5. Click Build knowledge index or Rebuild knowledge index; hybrid deployments show governed indexing controls. Wait for completion and review inaccessible/skipped sources.
FieldPurpose
KindLocal path or Web URL. Google Sheets URLs are routed through the authorized Sheets reader where supported.
ValueThe real local path or source URL. A teammate’s imported path needs local review.
LabelA recognizable source name, such as “Offers production procedure”.
CollectionControls domain scope used for retrieval; do not classify all specialized material as common.
Remove sourceRemoves the configured evidence source. Reconcile/rebuild as indicated; it does not delete the original document.
Search indexed sourcesFilters the source list in Settings, not a semantic question-answering prompt.

Collections shown in the app include Incident response & on-call, Production planning & delivery, Product & LiveOps, Game & content design, Tech Art production, Engineering & data, QA & quality, Localization, Operations & run intelligence and Common knowledge.

◇

Reference

Build, refresh and verify the index

Local mode embeds on the device; imported source definitions do not include an index. Ensure the embedding runtime/model is available. The local reader applies bounded file count, file size and web-fetch limits, so a large folder is not a promise that every file was indexed.

Reported valueHow to interpret it
ReadyThe currently checked retrieval contract is usable; still inspect freshness and source coverage.
Ready · degradedSome verified content may be retained while source access failed. Read the warning before relying on it.
Blocked / fail-closedThe required contract or freshness is not satisfied. Resolve the cause; do not call retained chunks fresh.
Indexed chunks / SourcesCoverage measurements, not a quality or certification score.
Last built / freshnessCompare with source changes. Rebuild after relevant content or skill revisions change.
Embedding contractHybrid retrieval binds provider, model/version, dimension and normalization. Changing these is a migration, not a cosmetic selector.

Ask a concrete question about a known procedure and inspect the cited source, collection and revision. If the answer lacks evidence, check source permissions and indexing before adding more prompt text. Never treat retrieved instructions as permission to bypass workflow rules.

◇

Reference

Configure Graphify and keep it current

  1. Set the intended project paths and inspect Platform Readiness → Code Intelligence. Distinguish Graphify installation, provider activation, current index and evaluation coverage.
  2. If the component is absent, use the supported setup assistant’s Graphify acceleration option when offered, or ask the maintainer for the supported installation/configuration.
  3. In Knowledge → Repository Intelligence & Agent Research, inspect the structural graph’s indexed revision versus repository HEAD.
  4. Choose Auto to refresh in the background when Git HEAD changes, or Manual for deliberate refreshes. Click Refresh graph now to build when available.
  5. Check completion time, node/edge counts, strategy and last error. A large graph from the wrong checkout or revision is not useful evidence for this run.

Graphify is the first Code Intelligence provider and remains experimental/advisory. Its local AST refresh is distinct from RAG embedding and from the Production Graph projection. Auto refresh follows HEAD changes; do not assume it captures arbitrary unsaved edits. Graph caches belong outside signed packages and are bound to source identity/revision.

◇

Reference

How to enrich agents effectively

Start with authoritative SOPs, field mappings, acceptance criteria and current examples in the appropriate RAG collections. Index the real code root for structural relationships, and retain explicit ticket/depot/asset identifiers so the Production Graph can connect evidence across domains.

For quality improvements, use representative labeled questions/tasks and measure whether retrieval finds the correct sources and explains conflicts. Add missing authoritative content and refresh stale indexes before increasing model size or graph traversal. Graphify installation alone is not proof that a consumer uses it; inspect provenance on the actual result.

◇

Reference

Production Graph and GraphRAG

The Production Graph can project Graphify, signed execution facts, item lineage, Jira/depot references, Unity GUID relationships and explicit artifact/build manifests. It distinguishes authoritative facts, derived relationships and hypotheses.

GraphRAG combines semantic/lexical retrieval with bounded graph expansion and reranking. Results must preserve source/revision provenance and report truncation. Missing graph data cannot invent missing semantic evidence. Projection replay rebuilds read models; it never reruns connector work.

Local storage uses SQLite; hosted adapters use PostgreSQL with workspace/game isolation. A implemented adapter or hermetic test is not proof that a particular hosted deployment is qualified. Production Graph contract.

◇

Reference

Executor knowledge crawl fields

FieldHow to choose it
Follow deep linksAllow bounded reading of task-related links from referenced wiki pages.
Max depth (hops)How far to follow links; zero reads only the referenced pages.
Max pagesHard cap on additional pages per draft/research pass.
Relevance strictnessRequired shared task terms; higher means fewer, more closely related links.
Save crawl settingsSaves within server-enforced bounds. Raising limits is not an authorization bypass.
◇

Reference

Local and hybrid deployment differences

Local: on-device retrieval and local state; local knowledge does not automatically sync to hosted storage. Hybrid: governed PostgreSQL/pgvector retrieval, explicit embedding contract and authorization filters; source changes enter governed ingestion and cutover remains operator-controlled.

Changing deployment mode requires the maintainer’s configuration and validation process. A setup bundle is not a database migration. RAG deployment modes.

◇

Reference

Automatic refresh after a run

Finalized run recording also requests a background RAG re-index through the knowledge change hook. This is separate from Graphify’s Git-HEAD refresh. Check the index status and last error: a refresh request does not prove every source was readable or successfully indexed.