Semantic Similarity

Semantic similarity measures conceptual likeness, not surface-level word matching. It uses sentence embeddings from models like Sentence-BERT or OpenAI's text-embedding models to map text into a vector space where proximity indicates related meaning.

• Example: The queries "automobile maintenance" and "how to service a car" have high semantic similarity despite sharing no keywords.
• This is critical for RAG, as user queries often use different vocabulary than the relevant source documents.

The primary mathematical operation is calculating the cosine similarity between two embedding vectors. This measures the cosine of the angle between them, providing a score from -1 (opposite) to +1 (identical).

• Normalized Vectors: Embeddings are typically L2-normalized, making cosine similarity computationally efficient as a dot product.
• Distance Metrics: Alternatives include Euclidean distance, but cosine similarity is dominant for text due to its focus on orientation over magnitude.
• Scores often range from 0 (unrelated) to 1 (equivalent), with relevant document-query pairs typically scoring above ~0.7.

Similarity scores are not absolute; they are relative to the embedding model used. Different models create different vector spaces.

• A score of 0.8 from one model (e.g., all-MiniLM-L6-v2) does not equate to the same perceived similarity as 0.8 from another (e.g., text-embedding-3-large).
• Thresholds must be calibrated per model and use case. The optimal threshold for determining a "relevant" document is found empirically via evaluation against labeled data.
• This characteristic necessitates consistent model usage throughout a pipeline's evaluation and production phases.

Semantic similarity is generally symmetric, but retrieval scenarios can be asymmetric. A short query embedding compared to a long document embedding may yield a different score than the reverse.

• Best Practice: Use a bi-encoder architecture trained for asymmetric retrieval (e.g., query and passage encoded separately but aligned).
• Pooling Strategies: For long documents, embeddings are often created by pooling (mean, max) sentence or chunk embeddings, which affects the similarity calculation.
• This asymmetry is why retrieval-specific embedding models outperform generic sentence transformers in RAG benchmarks.

It is the operational mechanism of dense retrieval. A vector database (e.g., Pinecone, Weaviate) indexes document embeddings. At query time, the query is embedded, and a k-nearest neighbors (kNN) search returns the documents with the highest similarity scores.

• This enables semantic search, finding documents that are topically related even without keyword matches.
• Performance is evaluated using metrics like Recall@K and NDCG@K, where K is the number of top results retrieved based on similarity score.
• The quality of the entire dense retrieval stage hinges on the semantic similarity metric's accuracy.

Beyond powering retrieval, semantic similarity is used as a direct evaluation metric. The average similarity score between a query and its ground-truth relevant documents is a strong indicator of embedding model and retrieval health.

• Monitoring: A drop in average query-document similarity over time can signal embedding drift or degradation in retrieval quality.
• A/B Testing: Used to compare the performance of different embedding models or chunking strategies.
• Limitation: It does not directly measure factual correctness (faithfulness) or answer quality, which require separate metrics like Answer Faithfulness or Grounding Score.

Cosine Similarity is the most common metric for calculating semantic similarity between two vector embeddings. It measures the cosine of the angle between two non-zero vectors in an inner product space, providing a value between -1 and 1.

• Calculation: Similarity = (A · B) / (||A|| ||B||). A value of 1 indicates identical orientation.
• Advantage: Efficient and invariant to vector magnitude, focusing solely on directional alignment.
• Use Case: The default scoring function for comparing query and document embeddings in vector databases.

Contrastive learning is the training paradigm used to create effective semantic similarity models. It teaches the model to pull similar items (positive pairs) closer in the embedding space while pushing dissimilar items (negative pairs) apart.

• Common Loss Functions: Multiple Negatives Ranking Loss (common for retrieval), Cosine Similarity Loss, Triplet Loss.
• Fine-Tuning Data: Requires labeled pairs of similar texts (e.g., (query, relevant document), (question, answer), (paraphrase1, paraphrase2)).
• Outcome: Produces an embedding space where cosine distance directly corresponds to semantic relatedness.

BERTScore is an automatic evaluation metric for text generation that computes similarity scores between a candidate text and one or more reference texts using contextual embeddings from models like BERT. Unlike traditional metrics based on lexical overlap (e.g., BLEU, ROUGE), it captures semantic meaning.

• Mechanism: It aligns each token in the candidate sentence with the most semantically similar token in the reference sentence using cosine similarity between their BERT embeddings, then computes a precision, recall, and F1 score based on these alignments.
• Use Case: Primarily used for evaluating machine translation, text summarization, and other generation tasks where semantic fidelity is more important than exact word matching.

Dense retrieval is a search paradigm where documents and queries are encoded into high-dimensional vector embeddings (dense vectors) by a neural network, typically a transformer-based bi-encoder. Retrieval is performed by finding documents whose embeddings have the highest semantic similarity (e.g., cosine similarity) to the query embedding.

• Contrast with Sparse Retrieval: Unlike traditional keyword-based (sparse) methods like BM25, dense retrieval matches based on conceptual meaning, enabling it to find relevant documents even when vocabulary differs.
• Foundation for Semantic Search: This technique is the foundational retrieval mechanism in most modern RAG architectures, directly relying on metrics of semantic similarity to rank passages.

Context relevance is a metric that assesses the degree to which the text passages retrieved and provided to a language model are pertinent and useful for answering a specific query. It evaluates the quality of the retrieval step in a RAG pipeline.

• Measurement: Often evaluated by having an LLM judge whether each retrieved passage contains necessary information to answer the query, or by calculating the semantic similarity between the query and the retrieved context.
• Impact on Generation: High context relevance is a prerequisite for answer faithfulness; irrelevant context increases the likelihood of the model hallucinating or generating a generic response.

Answer faithfulness (or factual consistency) is a metric that measures the extent to which a generated answer is factually consistent with and supported by the provided source context. It specifically targets hallucination within a RAG setting.

• Core Question: "Does the answer contain any statements that cannot be inferred from the given context?"
• Evaluation Method: Typically assessed using an LLM-based evaluator or by cross-referencing claims in the answer against the source documents. It is distinct from answer correctness, as faithfulness does not require the context itself to be factually true, only that the answer is loyal to it.

Reranking effectiveness refers to the improvement in retrieval quality achieved by applying a secondary, more precise neural ranking model to an initial set of candidate documents (often from a dense retriever). The reranker computes a more refined semantic similarity score.

• Two-Stage Process: A fast, recall-oriented retriever (e.g., using approximate nearest neighbors on embeddings) fetches a top-K set (e.g., 100 documents). A slower, cross-encoder model then re-scores this set by jointly processing the query and each document, providing a more accurate relevance score.
• Measured By: The lift in metrics like NDCG@K, MAP, or Precision@K after the reranking step compared to the initial retrieval.

A grounding score is a metric that evaluates the degree to which a model's generated output is substantiated by specific, attributable information from its provided source materials. It is a broader concept than faithfulness, often encompassing citation accuracy.

• Components: Can include answer faithfulness, source citation recall (were all used facts cited?), and source citation precision (were all citations accurate?).
• Operationalization: In evaluation frameworks, this may be implemented by prompting an LLM to extract all claims from an answer and verify their presence and attribution in the source context, producing a quantitative score.