Integration Capabilities
AI Fabrix provides a governed integration fabric for enterprise systems and data. Connectors run inside the customer’s Azure tenant, enforce Entra ID / RBAC, and respect metadata-aware retrieval so applications and agents only access data they are entitled to use.
API Integration
Scope
- REST/HTTP (JSON), GraphQL, and SOAP (legacy).
- OAuth2, API keys (stored in Azure Key Vault), mTLS where required.
- Pagination, rate-limit backoff, idempotency keys, and retry/circuit-breaker policies.
Capabilities
- Request signing, custom headers, HMAC verification.
- OpenAPI import for typed requests/responses and automatic schema validation.
- Webhook endpoints with verification secrets and replay protection.
- Fine-grained egress controls: allowlists per connector, per-environment quotas.
Database Integration
Supported patterns
- Federated read via service accounts with least privilege.
- Batch extract and CDC (change data capture) where sources support it.
- Caching via Redis; vectorization into PostgreSQL/pgvector when required.
Common engines
- Azure SQL / SQL Server, PostgreSQL, MySQL, Oracle.
- Data warehouses (read patterns): Azure Synapse, Snowflake, BigQuery (via APIs).
Controls
- Connection pooling, read-only roles, masked views for PII/PHI.
- Network isolation using Private Endpoints or self-hosted runners.
File System Integration
Sources
- SharePoint, OneDrive, Teams Files, Azure Blob/Files; SFTP for legacy file drops.
Document processing
- Incremental crawls based on modified/created timestamps.
- File-type filters (.docx, .pptx, .pdf, .txt, .xlsx, .csv, etc.).
- Inherited folder metadata (e.g., Deal Name, HubSpot ID) merged into file records.
- OCR where needed; text/metadata extraction pipelines with retry and dead-letter queues.
Security
- Source-permission checks on retrieval; metadata tags drive access control in RAG.
Cloud Service Integration
Azure-first
- Entra ID (SSO, SCIM), Key Vault, Storage, App Service/Container Apps, Front Door.
- Microsoft Graph for M365 (users, groups, files, calendars).
Other clouds
- AWS/GCP services accessed through API connectors or private routing.
- Consistent audit schema and egress policies across providers.
Third-Party Integrations
Business apps
- CRM/ERP/HR/Finance platforms (e.g., Dynamics 365, Salesforce, SAP, Workday, NetSuite, HubSpot).
- ITSM/Support tools (ServiceNow, Jira, Zendesk).
- Collaboration (Teams, Slack) for message/file ingestion under RBAC.
Approach
- Server-side connectors (no browser tokens) with typed inputs/outputs.
- Per-connector policy packs: rate limits, payload size caps, PII scrubbing rules.
- Observability: per-call logs, correlation IDs, metrics (latency, throughput, errors).
Custom Connectors
SDK & Plugin Framework
- Clean, developer-friendly SDK with dynamic input fields, dependent pickers, and output schemas.
- Safe execution model (containerized, server-side only); no credentials in code—Key Vault required.
- Versioned manifests, semantic versioning, and canary rollout support.
Developer experience
- Local test harness and contract tests.
- CI/CD templates for build, sign, scan, and publish to internal catalog.
- Linting for policy compliance (egress, secrets, PII handling).
Integration Patterns
- Event-driven: webhooks/queues trigger Flowise workflows; sub-second SLA where supported.
- Batch: scheduled extractions, transformations, and vectorization.
- CDC: stream updates into search/vector stores; reconcile with source-of-truth keys.
- Federated read (“zero-copy”): query on demand with source permissions; cache where allowed.
- RAG with metadata filters: retrieve only documents the caller is entitled to see.
- Human-in/on-the-loop: approvals, exception handling, and reversible actions.
Data Synchronization
Strategies
- Full load → incremental: initial backfill, then delta by timestamp or CDC.
- Key-based upsert: stable IDs map to source objects; conflict resolution via version/ETag.
- Schema evolution: additive fields handled automatically; breaking changes flagged in CI.
Quality & lineage
- Checksums and record counts per batch; drift detection vs. source.
- End-to-end lineage: source → transform → index/vector → consumer workflow.
- Rollback plans and replayable jobs from durable queues.
Governance
- Data minimization and retention policies per dataset.
- PII/PHI tagging, masking, and allow/deny lists enforced at connector and query-time.
