Search

Search IconIcon to open search

Mixture Of Encoders

Last updatedUpdated: by Jakub Žovák · 2 min read

Properties
created 04.12.2025, 15:12
modified 06.07.2026, 10:01
published Empty
sources Cursor Answer From
topics SuperLinked
authors Empty
ai-assisted No

# Papers

# How to Prevent Dwarfing By Large Semnatic Vector

  • To make a concatenated vector schema function correctly, practitioners use several normalization and structural strategies:
  1. Independent Space Normalization (Unit Length Scaling)
    • Before concatenation, you must normalize each sub-vector individually to a unit length of 1 (\(\Vert{}V\Vert{} = 1\)) Source.
      • The Result: If \(V_{text}\) is normalized to 1, and \(V_{metadata}\) is normalized to 1, they carry exactly equal weight in the final distance calculation, regardless of how many dimensions each possesses.
  2. Explicit Sub-Vector Weighting
    • Once both sub-vectors are normalized to a magnitude of 1, you can apply an explicit scalar weight (\(w\)) to each component prior to concatenation:
      $$V_{composite} = [ w_1 \cdot V_{text} \,\Vert{}\, w_2 \cdot V_{metadata} ]$$
    • If your business rules dictate that metadata constraints (like a price ceiling or a specific category) are highly critical, you can set \(w_2 = 2.0\) and \(w_1 = 0.5\). This ensures the metadata block dictates the geometry of the vector space more than the text chunk.

# Why Concatenation Works:

  • Preserves all information from each embedding space
  • No parameter learning required (simpler, more robust)
  • Linear separability: Different dimensions handle different semantic aspects
  • Query-time flexibility: Weighting happens during search, not storage
  • Scalable: Works efficiently with vector databases

# The Trade-offs:

  • According to research, there are more sophisticated methods, but they come with costs:
  • Simple concatenation (Superlinked’s approach):
    • ✅ Fast, no training needed
    • ✅ All information preserved
    • ✅ Query-time control via weighting
    • ❌ Treats all dimensions equally at storage time
    • ❌ No learned interactions between spaces

# Different Fusion Strategies

  1. Early Fusion (concatenation) - Combine before processing
  2. Late Fusion (weighted averaging) - Combine after separate processing
  3. Attention-based Fusion - Dynamic weighting based on context
  4. Bilinear Pooling - Captures pairwise interactions
  5. Gated Fusion - Selective information flow (LSTM/GRU-style)