@inferagraph/redis

@inferagraph/redis

Redis-backed RAG primitives for @inferagraph/core's AIEngine. One package, four pluggable stores plus a one-shot index provisioner:

Concept	Class	Factory
LLM response cache	RedisCacheProvider	redisCacheProvider
Multi-turn chat memory	RedisConversationStore	redisConversationStore
Vector embedding storage (RediSearch HNSW)	RedisVectorEmbeddingStore	redisVectorEmbeddingStore
Inferred-edge overlay (RediSearch HNSW)	RedisInferredEdgeStore	redisInferredEdgeStore
One-time index bootstrap	—	provisionRedisVectorIndex

The vector and inferred-edge stores require Redis Stack (RediSearch + RedisJSON). The cache and conversation stores work against any Redis 6+.

Installation

pnpm add @inferagraph/redis @inferagraph/core redis

Migrating from @inferagraph/redis-cache-provider

pnpm remove @inferagraph/redis-cache-provider
pnpm add @inferagraph/redis

The class names did not change. The legacy factory redisCache is retained as a deprecated alias of redisCacheProvider. The cache CacheProvider shape now also exposes delete(key) and accepts a per-call { ttlSeconds } on set (per @inferagraph/core@0.9.0).

Cache provider

import { AIEngine } from '@inferagraph/core';
import { redisCacheProvider } from '@inferagraph/redis';

const cache = redisCacheProvider({
  url: process.env.REDIS_URL ?? 'redis://localhost:6379',
  // Optional. If both maxEntries and ttl are unset, defaults to (500, '24h').
  maxEntries: 1000,
  ttl: '12h',
});

const engine = new AIEngine({ /* ... */ });
engine.setCache(cache);

// Wider 0.9.0 surface:
await cache.set('key', 'value', { ttlSeconds: 30 }); // per-call TTL wins
await cache.delete('key');                            // single-key delete
await cache.clear();                                  // SCAN+DEL by prefix (NOT FLUSHDB)

Option	Required	Description
url	one of url / client	Redis connection URL (redis://... or rediss://...).
client	one of url / client	Pre-built node-redis client. When provided, url is ignored.
prefix	No	Key prefix (default infera:cache:).
maxEntries	No	Maximum entries to retain. -1 disables. See defaults note below.
ttl	No	Construction-time TTL. Number (ms) or duration (5m, 2h, 7d, 1w). -1 / '-1' disables.

Defaults: when both maxEntries and ttl are unset, the provider defaults to (500, '24h'). Per-call { ttlSeconds } on set() always wins over the construction-time default.

Conversation store

import { AIEngine } from '@inferagraph/core';
import { redisConversationStore } from '@inferagraph/redis';

const conversations = redisConversationStore({
  url: process.env.REDIS_URL,
  // Optional defaults shown:
  keyPrefix: 'inferagraph:conversation',
  ttlSeconds: 86_400,
});

const engine = new AIEngine({ /* ... */ });
engine.setConversationStore(conversations);

Storage layout: one Redis LIST per conversation, keyed by <keyPrefix>:<conversationId>. appendTurn LPUSH-es a JSON-serialized turn, LTRIM 0 999 caps the per-conversation history at 1000 turns, and EXPIRE <ttlSeconds> refreshes the TTL on every append. Malformed entries are skipped with a console.warn rather than thrown.

Vector embedding store (RediSearch)

import { redisVectorEmbeddingStore, provisionRedisVectorIndex } from '@inferagraph/redis';

// Once at deploy time — idempotent. Creates BOTH the embeddings index AND the
// inferred_edges index by default; pass alsoProvisionInferredEdges:false to opt out.
await provisionRedisVectorIndex({
  url: process.env.REDIS_URL,
  embeddingDimensions: 3072,           // matches text-embedding-3-large
});

const embeddings = redisVectorEmbeddingStore({
  url: process.env.REDIS_URL,
  embeddingDimensions: 3072,
});

// engine.setEmbeddingStore(embeddings);

Storage layout:

• <keyPrefix>:embedding:<nodeId> — HASH containing nodeId, embedding (binary Float32 array), embeddingHash, embeddingModel, embeddingVersion, embeddingGeneratedAt.
• <keyPrefix>:embeddings:idx — RediSearch index over the hashes (HNSW vector field, COSINE distance).

searchVector(query, { top }) issues FT.SEARCH <idx> "*=>[KNN $top @embedding $vec AS score]" and converts the returned distance to similarity (1 - distance) so the contract's "higher = more similar" holds.

Inferred-edge store (RediSearch)

import { redisInferredEdgeStore } from '@inferagraph/redis';

const inferred = redisInferredEdgeStore({
  url: process.env.REDIS_URL,
  embeddingDimensions: 3072,
});

// aiEngine.setInferredEdgeStore(inferred);

Storage layout:

• <keyPrefix>:inferred_edge:<sourceId>:<targetId>:<type> — HASH with id, sourceId, targetId, type, score, sources, reasoning, perSource, optional embedding.
• <keyPrefix>:inferred_edges:idx — RediSearch index with TAG indexes on sourceId, targetId, type plus an HNSW vector index on embedding. getAllForNode(nodeId) queries by (@sourceId:{nodeId}) | (@targetId:{nodeId}) so both directions hit at O(log n).

set(edges) SCAN-deletes all existing <prefix>:inferred_edge:* keys, then HSETs the new ones (bulk replace, per the InferredEdgeStore contract). searchInferredEdges(query, top) issues a KNN query against the same index.

Provisioning

import { provisionRedisVectorIndex } from '@inferagraph/redis';

await provisionRedisVectorIndex({
  url: process.env.REDIS_URL,
  keyPrefix: 'inferagraph',                   // default
  embeddingDimensions: 3072,                  // default
  alsoProvisionInferredEdges: true,           // default
});

Idempotent: catches RediSearch's "Index already exists" and treats it as a no-op. Run on every deploy without fear of failures.

Notes

• All four stores connect lazily on first operation, so constructing them is cheap and won't throw.
• The package targets the protocol (Redis 6+ with RediSearch for vector stores), not a specific provider — the local redis://localhost:6379 and managed Redis Cloud / Elasticache / Upstash all work transparently.
• clear() on the cache uses SCAN (never KEYS), so it is safe on large datasets and never touches keys outside the configured prefix.

License

MIT