Implement Similarity Search Engine (Cosine / Dot Product)

[ustyuzhaninky]

Implement Similarity Search Engine (Cosine / Dot Product)

Labels: stage-8, backend, retrieval, embeddings, math

Description

Add a retrieval search module capable of performing k-nearest-neighbors lookups using cosine similarity or dot product over embedding vectors stored in VectorStore.

Requirements

• Implement a RetrievalEngine class with:
  • search(query_embedding, k=5)
  • returns top-k closest matches
• Similarity options:
  • cosine similarity (default)
  • dot product (optional)
• Convert stored embeddings and query embedding into numpy arrays or lists
• Ensure dimensionality matches
• Include ranking scores in output

Acceptance Criteria

• RetrievalEngine implemented with cosine similarity
• Returns sorted list of (id, score, text, metadata)
• Protect against empty index (returns empty list or error)
• Test accuracy with known vector inputs
• Edge-case tests for different vector sizes

References

Parent: Stage 8 — Simple Retrieval Queries