1
export class SemanticCache {
2
  private redis: RedisClientType;
3
  private bedrock: BedrockRuntimeClient;
4
  private configs: Record<string, CacheConfig>;
5

6
  // Tìm cache hit
7
  async get(prompt: string, configKey = "claude-sonnet") {
8
    const embedding = await this.embedPrompt(prompt);
9

10
    const results = await this.redis.ft.search("idx:llm-cache",
11
      `@modelId:{${config}} @embedding:[VECTOR_RANGE 0.08 $vec]=>{$YIELD_DISTANCE:1}`,
12
      { PARAMS: { vec: vectorToString(embedding) }, SORTBY: "VECTOR_DISTANCE", LIMIT: {from:0, size:1} }
13
    );
14

15
    if (!results.total) return null;
16

17
    // Check similarity threshold + TTL
18
    const similarity = 1 - parseFloat(results.documents[0].value.distance);
19
    if (similarity < threshold || age > maxAge) return null;
20

21
    return { response, hitCount, similarity };
22
  }
23

24
  // Embedding dùng Amazon Titan Embeddings
25
  private async embedPrompt(prompt: string): Promise<number[]> {
26
    const result = await this.bedrock.send(new InvokeModelCommand({
27
      modelId: "amazon.titan-embed-text-v2:0",
28
      body: JSON.stringify({ inputText: prompt, dimensions: 1024, normalize: true }),
29
    }));
30
    return JSON.parse(new TextDecoder().decode(result.body)).embedding;
31
  }
32
}