Embedding Model

The dominant architecture for retrieval, where queries and documents are encoded independently into dense vectors. This separation allows documents to be pre-computed and indexed offline, enabling fast Approximate Nearest Neighbor (ANN) lookups at query time. The core trade-off is speed versus depth of interaction: because the query and document never interact directly within the model, bi-encoders can miss subtle contextual relationships that a cross-encoder would catch.

• Key Models: Sentence-BERT (SBERT), Dense Passage Retrieval (DPR)
• Strength: Massive scalability via pre-computation
• Weakness: Lower precision on nuanced queries compared to cross-encoders

Modern embedding models are trained using contrastive loss functions, which explicitly teach the model to pull semantically similar pairs close together in vector space while pushing dissimilar pairs apart. The most common approach uses in-batch negative sampling, where other items in the same training batch serve as negative examples, dramatically improving training efficiency.

• Process: Anchor → Positive (similar) vs. Negative (dissimilar)
• Benefit: Produces highly discriminative representations
• Example: The all-MiniLM-L6-v2 model was trained with over 1 billion contrastive sentence pairs

Before embedding, raw text is segmented into tokens—sub-word units like 'play', '##ing'—using a tokenizer such as WordPiece or Byte-Pair Encoding (BPE). The model has a fixed context window (e.g., 512 tokens for many SBERT models), which defines the maximum input length. Text exceeding this limit must be truncated or chunked, making the window size a critical constraint for semantic completeness.

• Impact: Determines maximum chunk size in retrieval pipelines
• Trade-off: Larger windows capture more context but increase compute cost quadratically
• Example: text-embedding-3-large supports an 8192-token context window

Transformer models produce a vector for every input token, but retrieval requires a single, fixed-length sentence embedding. A pooling layer collapses these token-level outputs into one vector. Mean pooling averages all token vectors, capturing the overall semantic signal, while CLS pooling uses only the first special token's representation. The choice of pooling significantly impacts the quality of the resulting embedding.

• Mean Pooling: Smoother, more robust representation; standard for SBERT
• CLS Pooling: Relies on a single token to encode the entire sequence
• Max Pooling: Selects the maximum value per dimension across all tokens

The embedding dimension—the length of the output vector—represents a direct trade-off between semantic fidelity and storage cost. Higher dimensions (e.g., 3072) capture more nuanced features but require more memory and slower search. Matryoshka Representation Learning (MRL) allows a single model to produce embeddings at multiple resolutions, enabling users to truncate vectors to a smaller size (e.g., 256d) without catastrophic loss of accuracy.

• Storage: 1536 dimensions × 4 bytes = 6KB per vector
• MRL Benefit: One model serves multiple latency/cost profiles
• Example: OpenAI's text-embedding-3 models natively support MRL

General-purpose embedding models often underperform on specialized jargon, such as legal, medical, or financial text. Domain adaptation involves fine-tuning a pre-trained model like BERT-base on a corpus of domain-specific sentence pairs using a contrastive objective. This process re-calibrates the vector space so that domain-relevant semantic relationships are prioritized.

• Technique: Continued pre-training on domain corpora, then contrastive fine-tuning
• Data: Requires curated pairs of similar/dissimilar domain sentences
• Result: Significant boost in retrieval precision for specialized vocabularies

The fixed length of the output vector generated by an embedding model. Common dimensions range from 384 (lightweight) to 4096 (high-fidelity).

• Higher dimensions capture more nuanced semantic features but increase storage and latency
• Lower dimensions are computationally efficient but may lose fine-grained distinctions
• MTEB leaderboard benchmarks models across dimensions to guide selection
• Dimensionality directly impacts ANN index memory footprint and query speed

The standard metric for comparing two embedding vectors in dense retrieval. It measures the cosine of the angle between vectors, normalizing for magnitude.

• Score ranges from -1 (opposite meaning) to 1 (identical meaning)
• Preferred over Euclidean distance because it ignores vector length differences
• Computed as the dot product of L2-normalized vectors
• Underpins all semantic search ranking in vector databases

The preprocessing step that segments raw text into tokens—the atomic numerical units consumed by embedding models. Tokenization strategy directly affects model input quality.

• WordPiece, BPE, and SentencePiece are common algorithms
• Out-of-vocabulary words are split into subword tokens
• Token count determines whether text fits within the model's context window
• Mismatched tokenizers between chunking and embedding cause truncation errors

The maximum sequence length of tokens an embedding model can process in a single forward pass. This hard limit defines the upper boundary for chunk size.

• Models like text-embedding-3-large support up to 8,191 tokens
• Exceeding the window causes truncation, losing semantic information
• Longer context windows enable larger chunks but increase latency
• Directly constrains chunking strategy design decisions