Home · Work

Systems Work · No. 21

Enterprise Knowledge and Context Layer

Rajesh Ranjan Mahapatra

The shared context backbone every agentic use case at a large regulated enterprise stands on, from governed ingestion through intent-routed retrieval to a context envelope with full traceability.

Start reading
Systems WorkCase study

© 2026 Rajesh Ranjan Mahapatra. 6 min read

Enterprise Knowledge and Context LayerAll Systems Work
Home · Work

Enterprise Knowledge and Context Layer

The shared context backbone every agentic use case at a large regulated enterprise stands on, from governed ingestion through intent-routed retrieval to a context envelope with full traceability.

Challenge

Every flagship agentic project at a large regulated enterprise, from front-line assistance to financial crime screening and credit adjudication, was about to build its own ingestion, its own retrieval and its own governance.

Architectural move

Rather than argue from preference, five candidate architectures were scored against nine signals using structured multi-criteria methods, ranked three independent ways, with a simulation sweep across the wider priority space.

Control boundary

Every knowledge object carries an owner, a version, provenance and entitlements, enforced before retrieval rather than checked after.

Outcome

The result is a funded platform layer that the enterprise's flagship agentic projects consume rather than rebuild, proven end to end by a six-week vertical slice and carried by an operating model and a regulator-ready assurance story.

FIG.00 // SYSTEM MAPSTORE / RECORDGATEAGENTDETERMINISTICSIGNAL / DATAL0 · INGESTGoverned ingestionSourcesystemsValidationEnrichmentApproval andpublicationL1 · STORESFour projections, onetruthGraph:relationshipsVectorindex:meaningContentstore:payloads,lineageCatalogue:owner,versionL2 · RETRIEVEIntent-routedretrievalDeterministiclookupStructuredquerySemanticsearchGraphtraversalL3 · ASSEMBLEContext assemblyGovernedenvelopeRebuiltevery turnL4 · REASONReasoning agentsEvery usecaseconsumesNonerebuildsKNOWLEDGE OBJECTSBY INTENTRETRIEVED OBJECTSENVELOPETRACELineageunbroken fromsource torenderedanswerSYS/EKCL · FIG.00 · SCHEMATIC · NOT TO SCALE
FIG.00 // SYSTEM MAPSIGNAL / DATADETERMINISTICAGENTGATESTORE / RECORDL0 · INGESTGoverned ingestionSource systemsValidationEnrichmentApproval and publicationL1 · STORESFour projections, one truthGraph: relationshipsVector index: meaningContent store: payloads, lineageCatalogue: owner, versionL2 · RETRIEVEIntent-routed retrievalDeterministic lookupStructured querySemantic searchGraph traversalL3 · ASSEMBLEContext assemblyGoverned envelopeRebuilt every turnL4 · REASONReasoning agentsEvery use case consumesNone rebuildsKNOWLEDGE OBJECTSBY INTENTRETRIEVED OBJECTSENVELOPETRACELineage unbroken from source torendered answerSYS/EKCL · FIG.00 · SCHEMATIC
FIG.00 System map. One governed backbone: seven-stage ingestion, four stores, four retrieval modes and a per-turn envelope.

The problem

FIG.01 // THE PROBLEMREJECTED PATHGATEMODELSIGNAL / DATATODAYEvery project builds itsowningestion, retrieval andgovernance, per projectDECISION PAPEROne governed context backboneevery agentic use case consumes itrather than rebuilds itRESULTThe silos come backhard to defend to a regulator,does not scale→ REBUILT PER PROJECTTHE SILOS RETURN ←WHAT THE ESTATE LOOKED LIKEGAP · 01Knowledge not masteredcustomer knowledge content in manyplacesGAP · 02Tooling without guardrailsno prompt control for customer-facingcontentGAP · 03Ungoverned web contentpublic-website content outside centralgovernanceSYS/EKCL · FIG.01 · SCHEMATIC · NOT TO SCALE
FIG.01 // THE PROBLEMSIGNAL / DATAMODELGATEREJECTED PATHTODAYEvery project builds its owningestion, retrieval and governance, perproject↑ REBUILT PER PROJECTDECISION PAPEROne governed context backboneevery agentic use case consumes it ratherthan rebuilds it↓ THE SILOS RETURNRESULTThe silos come backhard to defend to a regulator, does notscaleWHAT THE ESTATE LOOKED LIKEGAP · 01Knowledge not masteredcustomer knowledge content in many placesGAP · 02Tooling without guardrailsno prompt control for customer-facing contentGAP · 03Ungoverned web contentpublic-website content outside centralgovernanceSYS/EKCL · FIG.01 · SCHEMATIC
FIG.01 The problem. Five flagship projects were about to build five context layers; the decision paper chose one.

Every flagship agentic project at a large regulated enterprise, from front-line assistance to financial crime screening and credit adjudication, was about to build its own ingestion, its own retrieval and its own governance. That does not scale, it is hard to defend to a regulator, and it rebuilds the very silos the programme existed to remove. The enterprise's customer knowledge content was not mastered in one place, the incumbent knowledge tooling offered neither the guardrails nor the prompt control needed to generate customer-facing content, and public-website content sat outside central governance entirely.

I wrote the architecture decision paper for the shared layer at CTO level and designed the platform it selected: a governed context backbone that every agentic use case consumes rather than rebuilds.

Why it was hard

FIG.02 // WHY IT WAS HARDSTORE / RECORDGATEMODELDETERMINISTICSIGNAL / DATAQAn agent questionone retrieval mechanismforced on everythingfails quietlyROUTEIntent routerchoose the mechanism thequestion wantsRELATIONSHIPGraph traversalwho is connected to whomRELATIONSHIPFACTKey lookupone exact answerFACTPOLICYSemantic searchwhat the text saysPOLICYSYS/EKCL · FIG.02 · SCHEMATIC · NOT TO SCALE
FIG.02 // WHY IT WAS HARDSIGNAL / DATADETERMINISTICMODELGATESTORE / RECORDQAn agent questionone retrieval mechanism forced on everythingfails quietlyROUTEIntent routerchoose the mechanism the question wantsRELATIONSHIPRELATIONSHIPGraph traversalwho is connected to whomFACTFACTKey lookupone exact answerPOLICYPOLICYSemantic searchwhat the text saysSYS/EKCL · FIG.02 · SCHEMATIC
FIG.02 Why it was hard. Knowledge is what an agent retrieves; context is how it decides what to retrieve. Different questions want different mechanisms.

Part of the confusion was linguistic. The word "context" was carrying four different meanings at once, from the documents an agent retrieves to the session state it accumulates, and teams were talking past each other while committing budgets. The design had to fix the vocabulary before it could fix the architecture: knowledge is what an agent retrieves; context is how it decides what to retrieve and how to process it.

The technical tension was between commitment and reversibility. A shared layer has to fix contracts early enough that five projects can build against it, while leaving components swappable, because the vector index or graph store chosen this year will not be the best one available in three years. And a single retrieval mechanism forced on everything fails quietly: relationship questions want a graph, factual lookups want a key, policy questions want semantic search, and a platform that offers only one of these pushes teams back into building their own.

Key decisions

FIG.03 // KEY DECISIONSDETERMINISTICOPTIONLED ACROSS THE SAMPLED WEIGHT SPACEGoverned backboneWON EVERY RANKINGSingle enterprise ontology:years of modelling firstPer-project builds: rebuild thesilosRent the whole layer: truth on avendor roadmapFIVE CANDIDATES · NINE SIGNALS · THREE RANKINGS · BUILD THE DURABLE, OWN THE TRUTH, GOVERNEVERY TURN, RENT THE RESTSYS/EKCL · FIG.03 · SCHEMATIC · NOT TO SCALE
FIG.03 // KEY DECISIONSDETERMINISTICGoverned backboneWON EVERY RANKINGSingle enterprise ontology: years ofmodelling firstPer-project builds: rebuild the silosRent the whole layer: truth on avendor roadmapFIVE CANDIDATES · NINE SIGNALS ·THREE RANKINGS · BUILD THE DURABLE,OWN THE TRUTH, GOVERN EVERY TURN,RENT THE RESTSYS/EKCL · FIG.03 · SCHEMATIC
FIG.03 Key decisions. Five architectures scored against nine signals, ranked three independent ways with a simulation sweep. Bar lengths are indicative.

Rather than argue from preference, five candidate architectures were scored against nine signals using structured multi-criteria methods, ranked three independent ways, with a simulation sweep across the wider priority space. One option, the governed backbone, led across the overwhelming majority of the sampled weight space and won every ranking. The losing options each failed instructively: a single enterprise ontology front-loads years of modelling before the first use case ships, per-project builds rebuild the silos, and renting the whole layer from a vendor puts the enterprise's most defensible asset, its own governed knowledge, on someone else's roadmap.

The core commitments were contracts, not products: a knowledge-object contract that every store honours, a small ontology instead of an enterprise one, intent-routed retrieval across four modes, and a context assembly service that returns a governed envelope. The guiding principle was blunt: build the durable, own the truth, govern every turn, rent the rest.

Architecture

FIG.04 // ARCHITECTUREREJECTED PATHSTORE / RECORDGATEAGENTDETERMINISTICSIGNAL / DATAP · 01-07Seven-stage pipelinesource → validate →enrich → approve →publishK · OBJGoverned knowledgeobjectslanding in four storesR · INTIntent-routedretrievalfour modesA · ENVContext assemblya governed envelopeAGTReasoning agentsconsume the envelopeTURN · NNext turn rebuiltfreshnothing accumulatedacross the sessionLOW CONFIDENCEQRNTQuarantinelow-confidence objectsare never loadedSYS/EKCL · FIG.04 · SCHEMATIC · NOT TO SCALE
FIG.04 // ARCHITECTURESIGNAL / DATADETERMINISTICAGENTGATESTORE / RECORDREJECTED PATHP · 01-07Seven-stage pipelinesource → validate → enrich → approve → publishK · OBJGoverned knowledge objectslanding in four storesR · INTIntent-routed retrievalfour modesA · ENVContext assemblya governed envelopeAGTReasoning agentsconsume the envelopeTURN · NNext turn rebuilt freshnothing accumulated across the sessionLOW CONFIDENCEQRNTQuarantinelow-confidence objects are never loadedSYS/EKCL · FIG.04 · SCHEMATIC
FIG.04 Architecture. Each turn is reconstructed from the envelope, so a contaminated retrieval dies at the end of the turn.

A single seven-stage pipeline carries content from source systems through validation, enrichment and approval to publication as governed knowledge objects. Objects land in four stores, each holding a different projection of the same truth: a graph for relationships, a vector index for meaning, a content store for payloads and lineage, and a catalogue for ownership and versioning. Retrieval is routed by intent across deterministic lookup, structured query, semantic search and graph traversal, and any low-confidence object is quarantined rather than loaded. The context assembly service composes what retrieval returns into a governed envelope that reasoning agents consume.

The context assembly service is deliberately reconstructive. Each turn is assembled fresh from the governed envelope rather than accumulated across a session, so a contaminated retrieval dies at the end of the turn instead of living quietly inside a conversation.

Trade-offs

FIG.05 // TRADE-OFFSREJECTED PATHGATEDETERMINISTIC▼ PAYSDESIGN CHOICE▲ BUYSSlower for the first projectT-01Shared layerPayback from the secondconsumer, compounding afterSemantic completenessT-02Small ontologyShips now; grows by accretionAssembly work every turnT-03Per-turn reconstructionContainment and replayabilityOperate moreT-04Rent lessA real operating modelSYS/EKCL · FIG.05 · SCHEMATIC · NOT TO SCALE
FIG.05 // TRADE-OFFSDETERMINISTICGATEREJECTED PATHT-01Shared layer▼ PAYSSlower for thefirst project▲ BUYSPayback from thesecond consumer,compoundingafterT-02Small ontology▼ PAYSSemanticcompleteness▲ BUYSShips now; growsby accretionT-03Per-turn reconstruction▼ PAYSAssembly workevery turn▲ BUYSContainment andreplayabilityT-04Rent less▼ PAYSOperate more▲ BUYSA real operatingmodelSYS/EKCL · FIG.05 · SCHEMATIC
FIG.05 Trade-offs. In a regulated setting, containment and replayability are worth more than the tokens.

A shared layer is slower for the first project than building its own pipeline, and the design accepts that openly; the payback arrives with the second consumer and compounds after that. The small ontology trades semantic completeness for shipping: it models only what the live use cases need, and grows by accretion. Per-turn reconstruction costs assembly work on every turn that an accumulating design would amortise, but it buys containment and replayability, and in a regulated setting those are worth more than the tokens. Renting less means operating more, and the platform carries a real operating model as a consequence.

Governance

FIG.06 // GOVERNANCESTORE / RECORDGATE / HUMANDETERMINISTICSIGNAL / DATALIN · 01Source systemLIN · 02Published objectowner · version ·provenance ·entitlementsLIN · 03Assembled envelopeentitlements enforcedbefore retrievalLIN · 04Rendered answera reviewer can see whereevery statement camefromLINEAGELINEAGELINEAGEONE STORYCustomer-facing agents and the internal estate consume the same layerone assurance story rather than severalSYS/EKCL · FIG.06 · SCHEMATIC · NOT TO SCALE
FIG.06 // GOVERNANCESIGNAL / DATADETERMINISTICGATE / HUMANSTORE / RECORDLIN · 01Source systemLINEAGELIN · 02Published objectowner · version · provenance · entitlementsLINEAGELIN · 03Assembled envelopeentitlements enforced before retrievalLINEAGELIN · 04Rendered answera reviewer can see where every statement camefromONE STORYCustomer-facing agents and theinternal estate consume the same layerone assurance story rather than severalSYS/EKCL · FIG.06 · SCHEMATIC
FIG.06 Governance. Lineage runs unbroken from source system to rendered answer.

Every knowledge object carries an owner, a version, provenance and entitlements, enforced before retrieval rather than checked after. Low-confidence objects are quarantined rather than loaded. Lineage runs unbroken from source system to published object to assembled envelope to rendered answer, which is what lets the enterprise show a reviewer exactly where any statement came from. Customer-facing agents and the internal knowledge estate consume the same governed layer, so there is one assurance story rather than several.

Outcome

FIG.07 // OUTCOMEDETERMINISTICSIGNAL / DATAVERTICAL SLICE6 weeksPROVEN END TO ENDPLATFORM LAYERbuilt per projectFundedCONSUMED, NOT REBUILTDECISION PAPER REUSED BEYOND THE PROGRAMMEScoring methodOption analysisCARRIED BY AN OPERATING MODEL AND A REGULATOR-READY ASSURANCE STORYSYS/EKCL · FIG.07 · SCHEMATIC · NOT TO SCALE
FIG.07 // OUTCOMESIGNAL / DATADETERMINISTICVERTICAL SLICE6 weeksPROVEN END TO ENDPLATFORM LAYERbuilt per projectFundedCONSUMED, NOT REBUILTDECISION PAPER REUSED BEYOND THEPROGRAMMEScoring methodOption analysisCARRIED BY AN OPERATING MODEL AND AREGULATOR-READY ASSURANCE STORYSYS/EKCL · FIG.07 · SCHEMATIC
FIG.07 Outcome. A funded platform layer that the flagship projects consume.

The result is a funded platform layer that the enterprise's flagship agentic projects consume rather than rebuild, proven end to end by a six-week vertical slice and carried by an operating model and a regulator-ready assurance story. The decision paper itself became a reusable artefact: the scoring method and the option analysis have been picked up beyond the original programme.

What I learned

FIG.08 // WHAT I LEARNEDSTORE / RECORDGATEDETERMINISTICAX-01Fix the vocabulary firstHalf the disagreement dissolved oncecontext stopped meaning four things.AX-02Scoring is not bureaucracyIt is how a decision survives thepeople who were not in the room.AX-03Never rent your truthThe most defensible asset is your owngoverned knowledge.SYS/EKCL · FIG.08 · SCHEMATIC · NOT TO SCALE
FIG.08 // WHAT I LEARNEDDETERMINISTICGATESTORE / RECORDAX-01Fix the vocabulary firstHalf the disagreement dissolved once contextstopped meaning four things.AX-02Scoring is not bureaucracyIt is how a decision survives the people whowere not in the room.AX-03Never rent your truthThe most defensible asset is your own governedknowledge.SYS/EKCL · FIG.08 · SCHEMATIC
FIG.08 What I learned. Three lessons from the engagement.

Fix the vocabulary before the architecture; half of the disagreement dissolved once knowledge and context stopped meaning four things. Structured option scoring is not bureaucracy, it is how you make an architecture decision survive the people who were not in the room. And the most defensible asset a regulated organisation owns in this space is not a model or a platform, it is its own governed knowledge, which is exactly the thing never to rent.