66 lines
1.9 KiB
Markdown
66 lines
1.9 KiB
Markdown
# RAGcore integration
|
|
|
|
## Objective
|
|
|
|
Use the existing central RAGcore project. MobilityOps must not implement embeddings, vector storage, chunking or its own answer-generation pipeline.
|
|
|
|
## Namespace
|
|
|
|
- tenant: `northstar-mobility-demo`
|
|
- workspace: `mobilityops`
|
|
- collection: `internal-procedures`
|
|
|
|
These values are configurable.
|
|
|
|
## Source documents
|
|
|
|
The ten Markdown files under `knowledge/procedures/` are authoritative PoC sources. Keep their IDs, versions and effective dates as metadata.
|
|
|
|
## Required adapter interface
|
|
|
|
```python
|
|
class KnowledgeProvider(Protocol):
|
|
async def health(self) -> KnowledgeHealth: ...
|
|
async def ask(self, question: str, actor: ActorContext) -> GroundedAnswer: ...
|
|
async def sync_manifest(self, documents: list[KnowledgeDocumentRef]) -> SyncResult: ...
|
|
```
|
|
|
|
Implement:
|
|
|
|
- `RAGcoreKnowledgeProvider`;
|
|
- `DemoKnowledgeProvider` using deterministic keyword/BM25-style local source retrieval only.
|
|
|
|
The demo provider is a resilience/test adapter, not a second RAG platform. It must return extracted source passages and a template summary; it must not pretend to be generative AI.
|
|
|
|
## Answer contract
|
|
|
|
```json
|
|
{
|
|
"answer": "...",
|
|
"evidence_state": "grounded | insufficient | unavailable",
|
|
"sources": [
|
|
{
|
|
"document_id": "damage-procedure",
|
|
"title": "Damage handling procedure",
|
|
"version": "1.3",
|
|
"section": "2. Immediate actions",
|
|
"excerpt": "..."
|
|
}
|
|
],
|
|
"provider": "ragcore",
|
|
"correlation_id": "..."
|
|
}
|
|
```
|
|
|
|
## Safety
|
|
|
|
- send actor scope and tenant/workspace with each request;
|
|
- enforce source allow-list for this PoC;
|
|
- never fall back to general model knowledge silently;
|
|
- no customer PII is indexed in RAGcore;
|
|
- log question metadata and source IDs, not unnecessary full prompts.
|
|
|
|
## Degraded mode
|
|
|
|
When RAGcore is unreachable, return `unavailable` and keep all operational functions available. When evidence is weak, return `insufficient` with the best source matches and no fabricated procedure.
|