OPEN-SOURCE ATLAS
The Enterprise AI Stack
Five complementary open-source intelligence domains — and the boundaries between them.
An AI-native business needs more than a language model — the architecture becomes useful only when provenance, authority and evaluation connect these capabilities into a coherent decision workflow.
Research & recommendation — not deployed Quantiviq services
Onyx — Document & Enterprise Knowledge Retrieval
GitHub ↗Onyx is the document-facing enterprise knowledge layer: search, retrieve and cite distributed organizational information. It helps answer what was documented and where. It is not the business metric source of truth, the temporal fact graph or an authorization engine.
What it answers
Permission-filtered documents and source IDs. → Citation-bearing documents/snippets with freshness/ACL metadata.
What it does not
Numerical ground truth over databases; autonomous approvals.
License & caveat
MIT (Community Edition). Onyx Community Edition is MIT with core chat/RAG/agents/actions; larger-enterprise controls may require Enterprise Edition — confirm feature-by-feature instead of calling every permission feature open-source.
Guard
Permission propagation must be tested in the deployed edition; avoid leaks.
See it in the Living Organization →
Boundary verified: 2026-10-26
One question, five layers
A fully synthetic walkthrough: "Revenue dropped after deployment; why?"
“Revenue dropped after deployment; why? (entirely synthetic)”
- Business question / owner intent
- Identity, data classification, source scope and consent
- ORCHESTRATOR (bounded, observable, read-first)
- Onyx: find approved PRD/policy/change notes, with citations
- OpenMetadata: verify Revenue metric + owner + lineage + freshness
- WrenAI: ask a certified metric question → safe query and aggregate
- Graphiti: what decisions/owners/changes were valid then?
- HolmesGPT: correlate deployment/alerts/logs → candidate RCA
- Evidence ledger → conflict check → uncertainty & alternatives
- Decision proposal (NOT an action)
- Human authority gate → approve / reject / request-more-evidence
- Simulated action, measured feedback, reversible learning
Layer comparison
| Layer | What it answers | What it does not | License & caveat | Inputs & outputs | Link |
|---|---|---|---|---|---|
| Onyx — Document & Enterprise Knowledge Retrieval | Citation-bearing documents/snippets with freshness/ACL metadata. | Numerical ground truth over databases; autonomous approvals. | MIT (Community Edition) | Permission-filtered documents and source IDs. → Citation-bearing documents/snippets with freshness/ACL metadata. | GitHub ↗ |
| OpenMetadata — Data Context, Meaning & Lineage | Certified metric definitions/owner/quality/freshness/lineage. | Generating uncontrolled SQL; serving a general chat model without metadata. | Apache-2.0 (upstream) | Catalog metadata, lineage, data contracts. → Certified metric definitions/owner/quality/freshness/lineage. | GitHub ↗ |
| WrenAI — Governed Analytics / Text-to-SQL | Reviewed SQL / aggregated results with query and metric provenance. | Unreviewed data definitions; public SQL access to proprietary data. | Apache-2.0 (OSS core) + commercial | Approved metric semantics + read-only scoped warehouse access. → Reviewed SQL / aggregated results with query and metric provenance. | GitHub ↗ |
| Graphiti — Temporal Relationships / Episodic Memory | Current/historical edges with source links. | Simple static catalog; a shortcut to full reasoning. | Apache-2.0 | Reviewed source episodes and canonical entities. → Current/historical edges with source links. | GitHub ↗ |
| HolmesGPT — Observability / Incident Intelligence | Evidence-backed incident hypothesis, confidence, verification plan. | General-purpose assistance; unsupervised remediation. | Apache-2.0; CNCF Sandbox (accepted 2025-10-08) | Read-only curated metrics/logs/events (synthetic demo). → Evidence-backed incident hypothesis, confidence, verification plan. | GitHub ↗ |