Relational + metric
Use deterministic queries for orders, inventory, billing, SLAs, governed measures, and other structured facts where precision matters.
Exact values · joins · aggregationsEnterprise AI context layer
Your agents know the model. Give them your operations.
Connect plant-floor reality and telecom network state with the business rules, customer impact, and evidence AI agents need to recommend—and safely execute—the next best action.
Not everything is semantic search
Route each question to the right kind of truth.
Embeddings help with meaning and similarity. They are the wrong tool for an exact alarm code, a governed KPI, a changing service topology, or a live order state. BodhiContextAI is being designed as a federated context plane that selects and combines the retrieval paths a task actually needs.
Keyword + lexical
Find exact part numbers, alarm codes, clauses, tickets, product IDs, and specialist terminology without fuzzy substitution.
Identifiers · BM25 · filtersSemantic + multimodal
Retrieve conceptually related manuals, notes, procedures, images, and unstructured knowledge when the wording does not match.
Embeddings · reranking · metadataGraph + topology
Traverse how assets, services, resources, customers, products, and dependencies connect across operational domains.
Entity resolution · paths · impactEvent + temporal
Reason over current state, event order, validity windows, changes, and prior decisions instead of treating every fact as timeless.
Streams · time series · memoryTool + action context
Call authoritative APIs when an answer must be checked live, then expose approved actions with identity, policy, and human gates.
MCP · OpenAPI · workflow toolsContext router
Classify the task → resolve identity and entitlements → choose one or more retrieval paths → rerank and reconcile evidence → return the smallest useful context package with provenance.
The 2025–26 architecture shift
Context is becoming operational infrastructure.
The frontier has moved beyond stuffing more text into a prompt. Production agents discover tools on demand, retrieve only task-relevant evidence, preserve useful memory, and cross an explicit policy boundary before they act.
Agents load the right context and tools when needed.
Large static prompts waste tokens and bury the useful signal. Current agent platforms increasingly discover tools, filter results, and compact working context as a task evolves.
Efficient context with MCP · Anthropic ↗MCP moves context; the enterprise still has to govern its meaning.
Open protocols reduce one-off agent integrations. The context plane still resolves identifiers, metrics, relationships, permissions, freshness, and which source is authoritative.
Model Context Protocol · Anthropic ↗Useful context includes what changed and what was learned.
Production agents need scoped memory for corrections, decisions, validity windows, and evolving operational state—not an unbounded transcript or a timeless knowledge graph.
Multi-layer data context + memory · OpenAI ↗The context boundary is becoming the action boundary.
The same layer that decides what an agent may see should constrain what it may do, record the evidence used, and route high-impact operations through policy and approval.
Trustworthy agent controls · Anthropic ↗The architecture
A context fabric, not another data copy.
The layer federates existing systems, chooses the right retrieval method, and packages the smallest useful evidence set for the task. Materialized indexes are selective; authoritative state stays authoritative.
Federated truth
SQL · lakehouse · metrics
Documents · tickets · media
Knowledge + topology graphs
Streams · telemetry · live APIs
BodhiContextAI
Context router + policy plane
Intent, entity + persona resolutionHybrid retrieval + graph traversalTemporal assembly + scoped memoryPolicy, provenance + evaluationTool contracts + approval gates
Context contracts
REST · GraphQL · SQL
MCP · OpenAPI tools
Events · agent-to-agent handoffs
Answers · actions · evidence
Persona-aware context
The same event should not produce the same view for everyone.
An operator, reliability engineer, plant manager, network engineer, care agent, and finance leader have different responsibilities and permissions. BodhiContextAI is being designed to assemble context and expose actions for the authenticated persona and the purpose of the task.
Who is asking?
Bind workforce or customer identity to tenant, role, team, site, region, entitlements, delegation, and session risk.
SSO · IAM · RBAC · ABACWhat may they know?
Apply row, field, document, entity, geography, and purpose restrictions before retrieval results reach the model.
Masking · filtering · purposeWhat helps this persona?
Select the vocabulary, depth, evidence, workflow history, and recommendations appropriate to the person’s job and current task.
Role view · preferences · memoryWhat may they do?
Separate read from action privileges, enforce tool allowlists and limits, and require human approval for high-impact changes.
Guardrails · approvals · least privilegeExample personaContext this role may receiveContext and actions denied
Product management
Approved, aggregated, and de-identified product signals; roadmap feedback; adoption trends.
Raw plant telemetry, network alarms, customer PII, detailed topology, and every operational control tool.
Plant operator
Assigned line state, active order, relevant SOPs, quality limits, and local shift history.
Other sites, unrestricted cost or HR data, and engineering or maintenance actions outside the operator’s authority.
NOC engineer
Authorized domains, alarms, topology, service impact, change windows, and runbooks.
Unmasked billing or customer data, unrelated regions, and network changes outside the approved control scope.
Customer care
Customer entitlement, service health summary, known incident, order state, and approved resolution steps.
Raw network configuration, sensitive topology, engineering credentials, and direct infrastructure controls.
Real-time context + delta intelligence
Fresh by contract. Reconstructable by design.
“Real time” is not one refresh rate. Machine signals may arrive in milliseconds, orders through change data capture, policies through versioned publication, and manuals on a scheduled sync. Each context product should declare its freshness objective and preserve the change history behind it.
Capture the change
Consume event streams, CDC logs, webhooks, API reads, and scheduled snapshots. Retain source offset, event time, ingestion time, schema version, and actor or service identity.
Kafka · CDC · OPC UA events · webhooksUpdate only what changed
Resolve impacted entities, calculate before-and-after deltas, then incrementally update relational views, search indexes, embeddings, graph edges, caches, and temporal memory.
Incremental index · graph delta · cache invalidationRe-evaluate access
Run policy checks when a record, relationship, entitlement, or classification changes. Revoke stale context and notify affected agents instead of waiting for the next query.
Policy version · entitlement delta · invalidationReplay decision time
Answer what the source said, which version the agent saw, what policy allowed, and what action followed at a specific point in time.
Valid time · system time · immutable audit ledgerChange pathSource event→Entity + policy delta→Affected projections→Context-changed event→Agent refresh or revocation
An evidence trail, not only a chat transcript.
Record the identity and purpose, policy decision, retrieval routes, source versions, context supplied, model and tool calls, approvals, action result, and downstream outcome. Sensitive payloads can be redacted while retaining a verifiable control trail.
- Trace IDs across agent, retrieval, and tool execution
- Policy decision and approval records
- Source freshness, lineage, and output citations
- Replayable evaluations and incident review
Learn safely without rewriting truth.
Capture explicit corrections, accepted or rejected recommendations, task outcomes, evaluator scores, and recurring retrieval failures. Promote useful learning into scoped, versioned memory only after validation.
- Personal memory for preferences and working style
- Role and team memory for reusable operating knowledge
- Curated enterprise memory for approved definitions and policy
- Offline evaluation gates, canary rollout, and rollback
Memory precedenceEnterprise policy and authoritative data→Approved team knowledge→Personal preferences→Session context
Lower scopes may personalize presentation and workflow; they cannot override permissions, governed definitions, safety controls, or authoritative operational state.
What we are building
Cloud native by design. Hosted on your terms.
We are building the BodhiContextAI enterprise context layer as modular, API-first cloud infrastructure. It is designed to connect to the systems you already run, use the model stack you choose, and fit the security boundary your organization requires.
Now buildingA governed context foundation for production agents
Pluggable retrievalUse the SQL, search, vector, graph, stream, and cache engines already approved.
Private connectivityConnect systems without exposing private data sources publicly.
Model + protocol choiceUse approved models and expose governed context through APIs or MCP.
Evidence envelopesCarry source, retrieval path, timestamp, policy, confidence, and lineage with every response.
Private networking is a standard enterprise pattern for agent workloads; for example, Microsoft documents agent deployments using customer virtual networks and private endpoints. View the architecture guidance ↗
01 · Manufacturing
Move from plant signals to coordinated action.
Production, quality, maintenance, engineering, and supply teams each see part of an exception. The context plane can combine OPC UA live state, Asset Administration Shell semantics, MES/ERP transactions, maintenance history, exact identifiers, manuals, and dependency graphs into one governed decision package.
Outcome story · Production exception
A line is losing throughput. What should happen next?
The agent connects the live machine signal with the active production order, recent engineering change, maintenance history, quality limits, material availability, and downstream schedule.
- DetectCycle time is drifting beyond the product-specific control range.
- ExplainA changed component and an overdue calibration are correlated with the deviation.
- EvaluateContinuing risks a quality hold; stopping now affects two downstream orders.
- CoordinateRecommend a controlled maintenance window, resequence orders, reserve the part, and open the quality check—with evidence and approvals attached.
Production exception resolution
Bring schedule, material, asset, labor, and process constraints into one decision instead of reconciling them in a war room.
MES · ERP · APS · WMSMaintenance in operating context
Prioritize work by failure risk, current product run, spare availability, technician skills, and downstream production impact.
IoT · EAM/CMMS · manuals · inventoryQuality and traceability
Trace a non-conformance through lots, suppliers, process parameters, inspection results, and engineering changes.
QMS · PLM · genealogy · supplier dataConnected planning
Surface how a demand, capacity, supplier, or maintenance change cascades through production, inventory, and customer commitments.
S&OP · planning · procurement · ordersStandards-aware direction: OPC UA supplies secure industrial information exchange from field to cloud; the Asset Administration Shell provides a standardized digital representation and semantic submodels for assets. OPC UA + AAS specification ↗ IDTA Release 26-01 ↗
02 · Telecom OSS + BSS
Connect network state to customer and revenue impact.
OSS knows what changed in the network. BSS knows who is affected and what was promised. The context plane can connect TM Forum-aligned service and resource entities, Open API tools, live alarms, topology, orders, SLAs, and a temporal decision graph so domain agents operate on the same service reality.
Outcome story · Service degradation
Thousands of alarms fire. Which one matters first?
The agent correlates RAN, transport, and core events with service topology, affected products, open orders, customer SLAs, recent complaints, planned work, and field capacity.
- CorrelateCollapse symptom alarms into one probable fiber-path failure.
- PrioritizeRank impact by critical services, enterprise SLAs, customer count, and revenue exposure.
- DecideCompare reroute, remote remediation, and dispatch options against capacity and policy.
- CoordinateOpen one incident, trigger the approved network action, brief care, notify affected customers, and preserve the evidence trail.
Service assurance and triage
Correlate alarms across domains, map them to services and customers, and prioritize remediation by business impact.
NMS · fault · performance · topologyOrder fallout and activation
Explain where an order stalled across catalog, orchestration, inventory, provisioning, billing, and partner dependencies.
CRM · order management · activation · billingCustomer care resolution
Give care agents the live service state, entitlement, device, interaction, and incident context needed for the next best action.
CRM · product catalog · SLA · knowledgeRevenue and capacity decisions
Connect usage, charging, leakage signals, demand growth, and network capacity to guide assurance and investment decisions.
Mediation · charging · billing · planningAI-native telecom direction: TM Forum catalyst work is combining domain agents, API-exposed OSS/BSS capabilities, intent-driven operations, and shared temporal knowledge graphs for cross-domain decisions. Temporal Knowledge Graph catalyst ↗ ODA + agentic network catalyst ↗
Start narrow. Compound value.
Prove one workflow, then turn its context into shared infrastructure.
We begin with the decision or action that matters, map the minimum context it needs, establish quality and governance, and expose it through a reusable contract. The next agent starts ahead.
Choose the decision
Define the user, workflow, systems, risk, and evidence that signal a useful outcome.
Build the context slice
Connect sources, resolve meaning, apply policy, and evaluate answers against real cases.
Productize the contract
Serve stable context interfaces, monitor quality, and extend across agents and business units.