Capability Matrix¶

A complete map of everything the Enterprise AI Fabric delivers, where to find it, and when to use it.

Full capability inventory¶

Capability	What it delivers	Primary location
LLM abstraction	Unified factory for OCI GenAI, OpenAI-compatible, and multi-modal providers	`inception_core/libs/src/llm/`
Multimodal processing	Image analysis, PDF extraction, video analysis utilities	`inception_core/libs/src/llm/`
OCI Guardrails	Content safety, topic filtering, and policy enforcement on all agent I/O	`inception_core/libs/src/metro/guardrails/`
Audit tracing (LangFuse)	Full request/response traces, token counts, latency, cost monitoring	`inception_core/libs/src/metro/auditing/`
Continuous evaluation (AE-09)	4-layer assertion/trajectory/groundedness/rubric scoring, CI-gate `evalFactory`	`inception_core/libs/src/metro/evals/`
Structured logging	OCI Logging integration, centralized log management	`inception_core/libs/src/metro/logging/`
Document Understanding	OCI Document Understanding key/value extraction as a direct-call tool	`inception_core/libs/src/tools/dus2_tool.py`
Agent pattern templates	7 canonical agentic patterns as runnable LangGraph examples	`inception_core/accelerators/patterns_agentic/`
Swarm multi-agent pattern	Peer-to-peer handoff topology — agents transfer control directly, no central router	`inception_core/accelerators/patterns_agentic/src/agent_swarm_pattern.py`
Memory-aware deep agents	Batteries-included harness: checkpointer + store + skills + tracing	`inception_core/accelerators/patterns_agentic/`
Short-term memory	LangGraph thread-based in-memory checkpointing	`accelerators/patterns_agentic/`
Long-term memory	Oracle ADB-backed `OracleDBSaver` checkpointer	`accelerators/patterns_agentic/`
Memory store	Oracle ADB-backed `OracleStore` virtual filesystem with vector search	`accelerators/patterns_agentic/`
Human-in-the-loop	Approval gate patterns before tool execution	`accelerators/patterns_agentic/`
End-user impersonation	Shared DB schema + session role table + VPD enforcement	`accelerators/patterns_solution/user_impersonation_*`
IAM identity propagation	IDCS SSO → API GW → FastAPI → MCP → DB token exchange	`accelerators/patterns_solution/identity_propagation_*`
A2A invocation	Agent-to-agent service call patterns	`accelerators/patterns_solution/a2a_invoke_services/`
Async Fusion agent invocation	Asynchronous invocation of Oracle Fusion Studio agents	`accelerators/patterns_solution/async_invoke_fusion_agents/`
IDCS browser login	Zero-dependency IDCS OAuth2 Authorization Code flow for a browser UI — server-side session, `HttpOnly` cookie, and the platform's empty-`groups`-claim fix	`accelerators/patterns_solution/idcs_browser_login/`
Reference chatbot	Multi-tool LangGraph chat agent (KB search / NL2SQL / web search, pluggable short- and long-term memory) for a recipe's conversational entry point	`accelerators/patterns_solution/reference_chatbot/`
SQLcl MCP server	SQL execution and schema introspection via FastMCP over HTTP	`inception_mcp_servers/src/mcp_sqlcl/`
Object Storage MCP server	OCI Object Storage namespace, upload, list, delete via MCP	`inception_mcp_servers/src/mcp_server_os/`
ADW MCP server	Autonomous Data Warehouse queries via MCP with token propagation	`inception_mcp_servers/src/mcp_adw/`
OCI Database MCP server	OCI Database Cloud Service control-plane ops (e.g. list Autonomous DB versions) via MCP with token propagation	`inception_mcp_servers/mcp_oci_control_plane_database/`
MCP registry pattern	Explicit tool exposure control via `TOOL_REGISTRY`	All MCP servers
OIDC / JWT MCP auth	Dual-mode authentication: interactive OIDC + token/JWT bearer	All MCP servers
Invoice Automation recipe	AP invoice classification, auto-pay, supplier analysis	`inception_recipes/invoice_automation/`
Smart Dispatch recipe	Multi-agent field service dispatch with HITL and Oracle memory	`inception_recipes/smart_dispatch/`
Meridian Bank recipe	Loan intake/underwriting swarm, deterministic decision boundary, dual-UI (customer + officer)	`inception_recipes/meridian_bank/`
EBS Invoice Reconciliation recipe	AP three-way match against a stubbed OIC/EBS boundary, live-validated end-to-end	`inception_recipes/ebs_invoice_reconciliation/`
Inception Skills	`recipe-builder` (use case → full recipe, plus a change-request workflow for amending one already built), `mcp-server-builder` (target system → MCP server), and `agentic-engineering` (AE-01..15 review) skills, spec-driven	`inception_skills/`
Production app scaffold	Layered project structure: API, CLI, agents, tools, config	`applications_stub/agentic_application/`
UI scaffold	Frontend scaffold for agentic app front-ends	`applications_stub/ui_application/`
Container packaging	Docker build/push scripts for OCIR deployment	`inception_devops/containers/`
Security Terraform	IAM, network, logging, and token-function IaC modules	`inception_devops/security/`

Capability groupings¶

Governance capabilities¶

These capabilities apply automatically to all agents. They are not opt-in — they are the baseline every deployment inherits.

Capability	What it enforces
OCI Guardrails	Content safety + topic policy on all agent I/O — before and after LLM calls
LangFuse	Full execution traces: prompts, tool calls, token counts, latency, cost — queryable per session
Continuous evaluation (AE-09)	Versioned eval suites (`metro.evals`) as a CI gate — golden/edge/adversarial/regression categories, each with its own pass-rate threshold
OCI Logging	Centralized structured log management with alerting rules and retention policies
IAM / OIDC	End-to-end identity propagation — no anonymous access at any layer
VPD	Row-level security enforced by Oracle at query time — policy lives in the DB, not in app code

Agent runtime capabilities¶

Core LangGraph + Oracle harness capabilities that every agent benefits from:

Capability	What it provides
LangGraph state machine	Conditional branching, cycles, and multi-step reasoning loops
OCI GenAI inference	Cohere, LLaMA, and other OCI-hosted models via the LLM factory
Thread checkpointing	Resume conversations across process restarts; replay or time-travel any session
Memory store	Virtual filesystem with keyword and vector search, shared across the agent fleet
Skills	Lazy-loaded, fleet-shareable tool bundles versioned in the memory store
Subagent fleet	Parallel subagents share a common `OracleStore` — intermediate results visible across agents
Swarm handoff	Peer agents transfer control to each other directly (`transfer_to_<agent>` tool → `Command(goto=..., graph=Command.PARENT)`) — no central router deciding what's next

Tool capabilities (MCP)¶

Server	Tools exposed
SQLcl MCP	`execute_sql`, `schema_information`, `test_connection`, `whoami`
Object Storage MCP	`get_namespace`, `upload_object_file`, `list_objects`, `delete_object`, `whoami`
ADW MCP	`query_adw`, `get_schema`, `test_connection`, `whoami` — with full IAM token exchange
OIC MCP	Oracle Integration Cloud workflow trigger — write-back channel to EBS

Security capabilities¶

Mechanism	Role in the security model
SSO via IDCS	Single sign-on; user identity established before any agent interaction
IAM group verification	Role gate checked by agent before any tool executes
IAM DB token exchange	Converts SSO token → short-lived DB token scoped to the end user; no shared credentials
VPD row-level security	Filters query results based on IAM role at the Oracle kernel level
EBS shielded via OIC	Agents never call EBS APIs directly; all EBS writes go through OIC workflows

Design characteristics¶

Characteristic	Detail
Modular	Capabilities are separated by domain and runtime concern; swap any layer independently
Composable	Recipes consume core building blocks and MCP services as drop-in dependencies
Cloud-native	OCI integrations (GenAI, IAM, ADW, Object Storage, Logging) are first-class citizens
Production-oriented	Includes test suites, settings modules, health checks, and operational documentation
Extensible	Add new tools by editing `registry.py`; add new patterns by following the catalog structure
Observable	Every layer emits traces, logs, and metrics to LangFuse and OCI Monitoring