Beyond Hybrid Search: Traversing Vector Spaces with Wormhole Vectors

Source: https://aiven.io/blog/beyond-hybrid-search-traversing-vector-spaces-with-wormhole-vectors
Author: Dima Kan (Aiven, Product Director of Search)

Summary

Dima Kan implements Trey Grainger’s Wormhole Vectors concept at Aiven, demonstrating how documents that bridge multiple retrieval spaces (sparse + dense + behavioral) can be used as traversal hubs for better retrieval than standard hybrid fusion.

The Problem with Standard Hybrid Search

Standard Hybrid Search (RRF fusion):

query → [sparse retrieval] → top-k₁
query → [dense retrieval]  → top-k₂
merge(top-k₁, top-k₂, via RRF) → final ranking

Limitation: The merge is a late operation. Each retrieval system still operates independently, potentially missing connections that only become visible when spaces are traversed jointly.

Example failure case: A relevant document might score 80th in sparse retrieval and 80th in dense retrieval — never making either top-k — but would be recognized as highly relevant if we could traverse between the spaces.

Wormhole Vectors in Practice

Kan’s implementation identifies “wormhole” documents:

• Documents that rank in the top percentile of both sparse AND dense retrieval for a given query cluster
• These documents act as hubs: their neighborhood in each space reveals different relevant documents

Query → [sparse top-100] ∩ [dense top-100] → wormhole candidates
       ↓
For each wormhole doc:
    expand its sparse neighbors
    expand its dense neighbors
    merge expanded neighborhoods
       ↓
Much richer candidate set than standard hybrid

Implementation at Aiven

Using PostgreSQL + pgvector + full-text search:

-- Find wormhole documents (high score in both sparse and dense)
WITH sparse_scores AS (
  SELECT id, ts_rank(search_vector, plainto_tsquery(query)) as sparse_score
  FROM documents
  ORDER BY sparse_score DESC LIMIT 100
),
dense_scores AS (
  SELECT id, 1 - (embedding <=> query_embedding) as dense_score
  FROM documents  
  ORDER BY dense_score DESC LIMIT 100
)
SELECT s.id, s.sparse_score, d.dense_score
FROM sparse_scores s JOIN dense_scores d ON s.id = d.id
-- These are wormhole documents

Results

Kan reports measurable improvements over standard RRF hybrid:

• Better recall for long-tail queries
• More diverse result sets
• Fewer “zero result” failures on edge cases

The improvement is most pronounced for queries where the relevant documents use different vocabulary than the query (vocabulary mismatch + semantic gap simultaneously).

Conceptual Framework

Wormhole vectors reveal that the retrieval space is not flat:

• Some documents are “attractors” in multiple spaces
• These attractors act as semantic transit hubs
• Traversing through attractors extends retrieval reach beyond direct similarity

Relation to Knowledge Graphs

The traversal metaphor connects to Knowledge Graph approaches:

• In KG search: traverse entity-relation-entity paths
• In wormhole search: traverse document-similarity-document paths across spaces

Related Articles

• Hybrid Search SPLADE Sparse Encoder — standard hybrid this extends
• SPLADE for Sparse Vector Search Explained — sparse component
• Nearest Neighbor Indexes for Similarity Search — dense component infrastructure

Related Concepts

• Wormhole Vectors — primary concept
• Hybrid Search — what this extends
• Dense Vector Retrieval — one space being traversed
• Sparse Vector Retrieval — another space
• Bag-of-Documents Model — related probabilistic framing

People

• Dima Kan — author, Aiven implementation
• Trey Grainger — Wormhole Vectors concept originator