Skip to content

Data Model and Tenant Isolation

Sudhanshu edited this page Nov 11, 2026 · 1 revision

Data Model and Tenant Isolation

Python request contract

RAGRequest(
    query="What are the Pro rate limits?",
    tenant_id="tenant_a",
    session_id=None,
    max_tokens=300,     # accepted but unused by extractive stub
    top_k=3,
)

Validation in RAGRequest.__post_init__ enforces query length 1–4000 characters after stripping, nonblank tenant_id, integer top_k between 1 and 20, and integer max_tokens between 1 and 2048.

Response contract

{
  "query": "original user query (not necessarily redacted)",
  "response": "Relevant knowledge-base excerpts:\n- ...",
  "sources": ["0:product"],
  "latency_ms": 12.34,
  "cached": false,
  "request_id": "random12chars"
}

The example numeric latency is illustrative, not a benchmark. sources entries concatenate Qdrant point IDs and category. The answer is not generated by an LLM.

Qdrant point

{
  "id": 0,
  "vector": "normalized BGE-small embedding, dimension inferred at runtime",
  "payload": {
    "tenant_id": "tenant_a",
    "category": "product",
    "text": "fictional knowledge-base passage"
  }
}

The built-in DOCUMENTS list contains four fictional points split between two tenants. Startup embeds all four and upserts into collection rag_demo_kb using cosine similarity.

Qdrant tenant filter

result = self.client.query_points(
    collection_name=COLLECTION,
    query=vec,
    query_filter=Filter(must=[FieldCondition(
        key="tenant_id", match=MatchValue(value=tenant_id))]),
    limit=top_k,
    with_payload=True,
    score_threshold=self.retrieval_min_score,
)

API mode determines tenant_id by matching the bearer token to TENANT_A_API_KEY or TENANT_B_API_KEY (configured via environment variables). The client cannot choose a tenant ID in /v1/query JSON. Direct Python calls accept any string tenant ID and do not authenticate a principal.

Cache-key isolation

Entries are only considered when all of the following match: tenant_id, session_id (including None), and top_k. Query text matches exactly or by normalized cosine similarity at/above the threshold. The response and sources are cached together.

This is demo-layer filtering, not proof of end-to-end tenant authorization, auditability, encryption or leakage prevention. See Security-and-Limitations.

Clone this wiki locally