Vector Search

Vector search is the core mechanism that enables semantic memory retrieval in PRX-Memory. Instead of matching keywords, vector search compares the mathematical similarity between query and memory embeddings to find conceptually related results.

How It Works

1. Query embedding: The recall query is sent to the configured embedding provider, producing a vector.
2. Similarity computation: The query vector is compared against all stored memory vectors using cosine similarity.
3. Scoring: Each memory receives a similarity score between -1.0 and 1.0 (higher is more similar).
4. Ranking: Results are sorted by score and combined with other signals (lexical match, importance, recency).

graph TB
    QUERY["Recall Query<br/>'SQL injection prevention'"]
    QEMBED["Query Embedding<br/>[0.12, -0.03, 0.87, ...]"]

    subgraph Storage["Stored Memory Vectors"]
        M1["Memory 1<br/>[0.11, -0.02, 0.85, ...]<br/>similarity: 0.97"]
        M2["Memory 2<br/>[0.45, 0.12, -0.33, ...]<br/>similarity: 0.23"]
        M3["Memory 3<br/>[0.09, -0.05, 0.82, ...]<br/>similarity: 0.94"]
    end

    RESULTS["Ranked Results<br/>1. Memory 1 (0.97)<br/>2. Memory 3 (0.94)<br/>3. Memory 2 (0.23)"]

    QUERY --> QEMBED
    QEMBED --> M1
    QEMBED --> M2
    QEMBED --> M3
    M1 --> RESULTS
    M2 --> RESULTS
    M3 --> RESULTS

Cosine Similarity

PRX-Memory uses cosine similarity as the distance metric. Cosine similarity measures the angle between two vectors, ignoring magnitude:

similarity(A, B) = (A . B) / (|A| * |B|)

Score	Meaning
0.95--1.0	Near-identical meaning
0.80--0.95	Highly related
0.60--0.80	Somewhat related
< 0.60	Likely unrelated

Combined Ranking

Vector similarity is one signal in PRX-Memory's multi-signal ranking. The final score combines:

Signal	Weight	Description
Vector similarity	High	Semantic relevance from embedding comparison
Lexical match	Medium	Keyword overlap between query and memory text
Importance score	Medium	User-assigned or system-computed importance
Recency	Low	More recent memories get a small boost

The exact weighting depends on the recall configuration and whether embeddings and reranking are enabled.

Performance

The 100k-entry benchmark shows:

Metric	Value
Dataset size	100,000 entries
p95 latency	122.683ms
Threshold	< 300ms
Method	Lexical + importance + recency (without network calls)

INFO

This benchmark measures the retrieval ranking path only, without network embedding or rerank calls. End-to-end latency depends on provider response times.

Scaling Considerations

Dataset Size	Recommended Approach
< 10,000	Brute-force cosine similarity (JSON or SQLite backend)
10,000--100,000	SQLite with in-memory vector scan
> 100,000	LanceDB with ANN indexing

For datasets exceeding 100,000 entries, enable the LanceDB backend for approximate nearest neighbor (ANN) search, which provides sub-linear query time.

Next Steps

• Embedding Engine -- How vectors are generated
• Reranking -- Second-stage precision improvement
• Storage Backends -- Choose the right storage backend