mirror of
https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools.git
synced 2026-06-17 23:09:35 +00:00
4.6 KiB
4.6 KiB
Technology radar — Tier-1 Dealix (official vs optional vs pilot)
Purpose: Separate what we ship today, what we commit to architecturally, and what stays a pilot/pattern until benchmarked. Do not treat pilot items as production dependencies without an ADR and evidence.
Execution criteria: execution-fabric.md. Spike policy: ../adr/0001-tier1-execution-policy-spikes.md.
Tier definitions
| Tier | Meaning |
|---|---|
| Official (now) | In repo, used in dev/staging/prod paths as implemented today |
| Official (target) | ADR-approved direction; may not be wired yet — document status honestly |
| Strong optional | Adopt when scale/integration needs justify ops cost |
| Pilot / pattern | Try behind flags; benchmark; no core coupling |
Radar table
| Technology | Tier | Role | Notes |
|---|---|---|---|
| FastAPI + PostgreSQL | Official (now) | API + operational store | Tenant isolation remains mandatory |
| Celery + Redis | Official (now) | Async tasks, schedules | See execution fabric for graduation criteria |
| LangGraph (where used) | Official (now) / partial | Stateful agents, HITL | Not a substitute for all long-running business workflows |
| Temporal | Official (target) | Crash-proof workflows, worker versioning | Planned — spike only until ADR passes |
| CloudEvents + JSON Schema + AsyncAPI | Official (target) | Event contracts | Discipline in events-and-schema.md |
| OpenTelemetry | Official (target) | Unified traces/metrics/logs | Incremental adoption |
| OPA / Rego | Strong optional / target | Policy PDP over JSON | See trust-fabric.md target table |
| OpenFGA or Cedar | Strong optional / target | Fine-grained authorization | Same |
| Vault (or cloud secret manager) | Strong optional / target | Secrets, dynamic creds | Rotation + audit |
| Keycloak (or enterprise IdP) | Strong optional / target | SSO / B2B identity | Map to customer IAM |
| Airbyte | Strong optional | Connector / ingestion plane | Reduces agent→vendor sprawl |
| Unstructured | Strong optional | Document extraction (PDFs, CIMs) | Pair with retention + S-class |
| dbt-style semantic layer | Strong optional | Single KPI source of truth | Avoid five definitions of “revenue” |
| Great Expectations | Strong optional | Data quality checkpoints | |
| OpenLineage or OpenMetadata | Strong optional (pick one first) | Lineage / catalog | Do not run two overlapping catalogs without cause |
| Neo4j | Pilot | Graph intelligence | Only when relationship reasoning is proven required |
| LangSmith | Strong optional | LangGraph/LangChain observability + evals | Commercial; evaluate vs OSS |
| Phoenix (Arize) | Strong optional | OTel-native tracing/evals | |
| Promptfoo | Strong optional | CI red-team / eval harness | |
| Guardrails AI | Strong optional | I/O validators | |
| ToolProof (concept) | Pilot / pattern | Verification ledger pattern | Not a hard dependency by default |
| MemPalace / similar | Pilot | Memory product | Benchmark before core |
| Flowise | Pilot | Internal sandbox only | Not core runtime |
| Local inference (generic) | Official (pattern) | Via adapter + health checks | No hardcoded vendor lock-in in agents |
External references (documentation)
Use upstream docs for detailed semantics (versions change). Prefer pinning versions in ADRs when adopting.
- OpenAI: Responses API, Agents SDK, structured outputs, MCP — OpenAI Platform documentation
- LangGraph: LangChain / LangGraph documentation
- Temporal: Temporal documentation
- CloudEvents: CloudEvents spec (CNCF)
- JSON Schema: json-schema.org
- AsyncAPI: asyncapi.com
- OpenTelemetry: opentelemetry.io/docs
- OPA: Open Policy Agent
- OpenFGA: openfga.dev
- Cedar (AWS): Cedar policy language
- GitHub: rulesets, environments, OIDC — GitHub Docs
- NIST AI RMF, OWASP LLM Top 10 — see saudi-compliance-and-ai-governance.md
Review
Revisit this radar quarterly or when adding a new external system. Update ../dealix-six-tracks.md status table when a pilot graduates to official.