ariffazil ⚒️ Redis + Qdrant Integration — Caching & Vector Memory
DITEMPA BUKAN DIBERI — Forged, Not Given.
Purpose
Redis as caching layer for MCP tool registry and session state; Qdrant as vector store for witness embeddings, semantic memory retrieval, and similarity search across federation knowledge.
When to Use
- Caching MCP tool lists, health probes, or computed results (Redis)
- Session state and lease TTL management (Redis)
- Semantic search across witness receipts, seals, or docs (Qdrant)
- Embedding storage for tri-witness evidence vectors (Qdrant)
- Tool/skill discovery via semantic similarity (Qdrant)
When NOT to Use
- Relational data that needs ACID — use
postgres-schema-design - Append-only immutable records — use
vault999-witness - File/blob storage — use S3/R2
Constitutional Floor Alignment
| Floor | Application |
|---|---|
| F1 AMANAH | Redis keys have TTL; stale cache never blocks fresh reads |
| F2 TRUTH | Cache headers (stale-while-revalidate) published with every cached response |
| F4 CLARITY | One cache key naming convention across all organs |
| F11 AUDIT | Cache misses and evictions logged; Qdrant writes carry provenance |
| F12 INJECTION | Vector embeddings sanitized before upsert — no prompt injection in stored vectors |
Commands & Patterns
// Redis — tool registry cache with TTL
const CACHE_TTL = 300; // 5 minutes
await redis.set(`tool:${organ}:list`, JSON.stringify(tools), { EX: CACHE_TTL });
// Redis — session state
await redis.set(`session:${sessionId}`, JSON.stringify(session), { EX: 3600 });
// Qdrant — witness collection upsert
const point: PointStruct = {
id: receiptId,
vector: embedding, // 1536-dim from embedding model
payload: {
witness_type: 'tri_witness',
organ: 'geox',
verdict: 'CONSENSUS',
timestamp: Date.now(),
},
};
await qdrant.upsert('witness_receipts', { points: [point] });
// Qdrant — semantic search
const results = await qdrant.search('witness_receipts', {
vector: queryEmbedding,
limit: 10,
with_payload: true,
});
Refusal Surface
- ❌ Redis persistence for critical audit data (use VAULT999)
- ❌ Qdrant as primary data store (vectors + relational in Postgres)
- ❌ Storing secrets or keys in Redis
- ❌ Embedding user PII into Qdrant vectors
- ❌ Cache-aside without invalidation strategy