+−⟲ AN IMAGINARY AI SYSTEM — REFINED ARCHITECTURE request flow • shared inference • agent coordination • retrieval • memory • tools • versioned knowledge SYSTEM-WIDE CONTROL PLANE ONLINE REQUEST + REASONING LOOP 1 · REQUEST + QUERY INTELLIGENCE 2 · ORCHESTRATION 3 · MULTI-AGENT RUNTIME 4 · EVIDENCE + RESPONSE SHARED MODEL GATEWAY SIDECAR ONLINE RETRIEVAL PLANE RUNTIME SERVICES AGENT SERVICE SHAFT MEMORY TOOLS + HOOKS OFFLINE INGESTION + ACTIVE KNOWLEDGE STORES ARROW SEMANTICS + READING GUIDE CONTROL BUS TO PIPELINES RUNTIME CONTROL dotted gray = control relation, not request flow DIRECT ROUTE result / revise SHARED MODEL CALL BUS SHARED MODEL RESULT BUS TOOL SERVICE CALL MEMORY RESULT TOOL RESULT FINAL MODEL CALL RETRIEVE ROUTE RETRIEVAL AGENT SERVICE CALL filters apply to every search lane CANDIDATE RESULT BUS RETRIEVED EVIDENCE + PROVENANCE SELECTIVE MEMORY WRITE SAFE / READ-ONLY BYPASS RISKY / IRREVERSIBLE TOOL RESULT COLLECTION BUS PARALLEL TOOL EXECUTION BUS INDEX READ DEPENDENCIES VERSIONED INDEX WRITE BUS · validated candidate → activate → rollback CALL RAIL RESULT RAIL SECURITY + GOVERNANCEidentity • tenant isolation • policyprivacy • data boundaries OBSERVABILITY + TRACINGrequest spans • model calls • toolsretrieval • memory • failures EVALS + FEEDBACKquality • groundedness • regressionreplay • human feedback COST + LATENCY + RESILIENCEbudgets • timeout • retrycircuit breaker • fallbacks CACHE + RATE LIMITSresponse • retrieval • embeddingquotas • backpressure VERSIONING + PROVENANCEmodels • prompts • schemas • chunksindexes • lineage • rollback ONLINE CONTROL PORT RETRIEVAL CONTROL PORT RUNTIME CONTROL PORT OFFLINE CONTROL PORT REQUESTtext • image • eventidentity + tenant INPUT GATEauth • safetyrate limit QUERY TRANSLATIONnormalize • rewrite • expanddecompose • languageintent • constraints • schema TASK CLASSIFIERdirect answerretrieveagent workflowtool action ROUTINGMODELpath • modelbudget • risk PLAN + STATEsteps • dependencieslimits • checkpointstop conditions SUPERVISORdelegatejoin • stop RESEARCHAGENTgatherevidence RETRIEVALAGENTsearchstrategy TOOLAGENTact onsystems MEMORYAGENTread /propose write CRITICAGENTverify /challenge SYNTHESISAGENTcombine resultsfor evidencemerge ALL AGENTSMODEL CALL / RESULT ROUTER MODEL CALL / RESULT MEMORY SERVICE CALL TOOL SERVICE CALL TASK BUS RESULT BUS DIRECT ROUTE AGENT RESULTS TOOL RESULTS MEMORY CONTEXT RETRIEVED EVIDENCE EVIDENCEMERGEmerge, nota decision CONTEXT ASSEMBLYevidence + agent resultsmemory + tool outputswindow budgetcitations + schema PROMPT PACKAGEsystem policyselected contextoutput contract MODELRESULT RESPONSESYNTHESISmodel result+ evidence RESPONSE GATEgrounding • policyschema • confidencecitations ANSWER /ACTIONcaller CALLIN MODEL POLICY ROUTERcapability • cost • latency • risk • provider health • fallback MODEL POOL SLMfast / cheap LLMdeepreasoning MULTIMODALvision / audio PROVIDER ADAPTERShosted • local • failover • capability normalization INFERENCE CONTROLStoken budget • timeout • retry • cache • safety MODEL CALL TELEMETRYusage • latency • safety events • trace correlation GATEWAY RESULTstructured output • tool request • refusal • model metadata RESULTOUT SHARED SERVICE — NOT A SEQUENTIAL STAGERouting, every agent, and final synthesis use the same call/result rails. SEARCH PLANNERqueries • filtersrecall strategy QUERY VARIANTSrewrite • expandhypothetical answer METADATA + ACLtenant • type • timepermissions SEMANTIC SEARCHvector similarityANN / exact BM25 SEARCHterms • fieldskeyword precision GRAPH SEARCHentities • relationspaths / communities HYBRID FUSIONnormalize scoresmerge • diversify DEDUP + DIVERSITYcollapse overlapsource balance RE-RANKERcross-encoder / LLMprecision ordering EVIDENCE PACKchunks • parent contextsource boundaries PROVENANCEcitations • scoresindex + chunk version MEMCALL MEMRESULT TOOLRESULT TOOLCALL MEMORYCALL MEMORY SEARCHsemantic • recencyimportance • scope WORKINGturn stateshort TTL PERSISTENTfacts • preferencesepisodes WRITE POLICYconsent • confidencecompact • decay • delete MEMORYRESULT reads are normal flow • writes are selective TOOLCALL TOOL REGISTRYcapability • schemaauth • risk class SELECT + EXECUTEvalidate argstimeout • retry APIsservices CODEsandbox DB / WEBread / act HOOKSbefore • afterretry • event HUMAN APPROVAL BOUNDARYrequired before risky or irreversible side effects TOOLRESULT ACTIVE STORES — READ PATH METADATACATALOGACL • lineageretention VECTORINDEXembeddingsANN settings BM25INDEXterms • fieldsanalyzers KNOWLEDGEGRAPHentities • edgescommunities DOCUMENT +CHUNK STOREsource textparent context INDEX + SCHEMAREGISTRYactive versionscompatibility INGESTION PIPELINE — WRITE PATH SOURCESdocs • web • DB • files • messages • events CONNECTORScrawl • subscribe • import • checkpoint PARSE + OCRextract structure • retain source offsets CLEAN + DEDUPEnormalize • classify • content hash CHUNKINGsemantic / structural • overlap • parent-child METADATA + ACLtenant • time • type • retention • access GRAPH ENGINEERINGentity / relation extraction • resolution • communities EMBEDDING GENERATIONmodel • dimensions • batch • version CANDIDATE INDEX BUILDupsert shadow version • snapshot • schema check VALIDATE + ACTIVATE GATEquality • dead-letter • retry • activate • rollback solid = request / data flow double = agent task ↔ result dashed purple = shared service dependency dashed teal = store / index read dotted gray = control relation The Model Gateway is a shared sidecar: the routing model, every agent, and final synthesis use explicit CALL and RESULT rails. Buses are named and junctions are marked; no connector is decorative.