 Zimin A.N.
|
e4cd7b9709
|
feat(ai): Phase 1 step 3-6 — LLM adapter + per-field AI suggest endpoint + UI
Approach C upgrade на templates foundation (!233). Admin может попросить AI
предложить новое поле — system prompt force'ит structural JSON Schema
fragment, admin review через accept/edit/reject. Bitemporal draft workflow
unchanged downstream.
## Backend
- `LlmAdapter` — OpenAI-compatible /v1/chat/completions client (Ollama,
vLLM, OpenAI). Bean conditional на ordinis.ai.enabled=true.
- `AiSchemaService` — system prompt + 3 few-shot examples, JSON output
parsing (с markdown fence stripping), validation (fieldName snake_case).
In-memory circuit breaker: 10 fails в 60s → open 5 min.
- `AiSchemaController`:
- GET /api/v1/ai/info → 200 (enabled) / 404 (disabled). Frontend probe.
- POST /api/v1/ai/suggest-field → suggestion. Per-IP rate limit 30/min.
- All bean-conditional: ordinis.ai.enabled=false → controllers нет →
endpoints 404 → frontend hides AI buttons automatically.
## Frontend
- `useAiFeatureAvailable` query — probes /ai/info, cache 5 min
- `useAiSuggestField` mutation — 30s timeout (LLM может быть slow)
- `AiFieldSuggestPanel` component — inline expandable: prompt → preview →
accept/retry/reject. Error handling per HTTP status (503/429/422/502/404).
- DictionaryEditorDialog: AI toggle visible на schema tab когда
aiAvailable=true. Accept wraps suggestion в mini-schema → parseSchemaJson
(reuse existing kind-inference) → push в properties (overwrite если
duplicate fieldName).
- i18n RU + EN
## Config
- application.yml: ordinis.ai.{enabled,endpoint,model,bearer-token,
timeout-seconds}
- Helm writer.yaml: ORDINIS_AI_* env vars, optional secretRef для bearer
- values.yaml: ai.{enabled,endpoint,model,timeoutSeconds,bearerTokenSecretRef}
defaults disabled
Pair с infra MR — values-staging enables AI с vortex.nstart.cloud Ollama
serving qwen2.5:7b.
|
2026-05-16 13:55:16 +03:00 |
|