Cosine Similarity

Semantic search converts text into high-dimensional vectors and finds the most relevant conclusions by computing cosine similarity, allowing queries phrased in different ways to match semantically equivalent results.

余弦相似度是衡量两个向量方向接近程度的指标，值域在-1到1之间，越接近1表示语义越相似

Cosine similarity has a value range of [-1, 1]; in semantic retrieval, attention is typically focused only on the positively correlated range of 0 to 1, with values closer to 1 indicating greater semantic similarity.

Cosine similarity only cares about vector direction and not length, with values strictly falling in the range [-1, 1].

When feature vectors are L2-normalized, cosine similarity is equivalent to the dot product operation.

Cosine similarity measures the cosine of the angle between two vectors, with the formula cos(θ) = (A · B) / (||A|| × ||B||).

Cosine similarity measures how close two vectors are in direction, with a value range of -1 to 1; the closer the value is to 1, the more similar the semantics, and because it is insensitive to vector magnitude and stable to compute, it has become the most mainstream choice in RAG scenarios.

Embedding models are deep neural networks that map text into a fixed-dimensional dense vector space (such as 768 dimensions or 1536 dimensions); cosine similarity is commonly used to measure semantic closeness, with a value range from -1 to 1.

Cosine similarity measures the angle between two vectors, where a result of 1 means the same direction (the same meaning) and 0 means irrelevant; it ignores vector length, so a short question can also match a long document.

The ingestion process of generative engines: the crawler parses HTML into fixed-token semantic chunks (usually 250–500 tokens); each chunk is converted into a high-dimensional vector; when the AI agent needs an answer, it selects the chunk with the highest cosine similarity and the densest entity coverage.

Similar Issue search works by generating a normalized sentence vector for each stored Issue, computing and ranking the cosine similarity between a new report and them, and returning the five most similar ones along with their original GitHub links.

The cosine similarity between two embedding vectors is meaningful only when they are produced by the same mapping function (the same model); changing the model (a new checkpoint, a new tokenizer, or a change in normalization) produces embeddings with the same shape but a different space.

There are three main ways to measure vector distance: cosine similarity, Euclidean distance (L2), and inner product (IP).

Embedding models map queries and documents into the same vector space and match them using measures such as cosine similarity; when a query uses domain abbreviations, internal codenames, or wording whose distribution differs greatly from the training corpus, the correct document may be ranked very far down or may not even appear in the candidate set.

Evaluation metrics based on semantic similarity (such as cosine similarity) cannot capture differences at the level of behavioral constraints and have systematic blind spots.

The conversion between cosine distance and similarity is: similarity = 1 − distance, where a distance of 0 means perfect agreement and a distance of 2 means complete opposition.

The semantic cache uses a cosine similarity threshold of 0.92 as the criterion for determining a cache hit.

A cosine similarity threshold is reliable only if it has been measured on your own corpus with the exact model in use, and re-measured whenever either changes.

If a shared direction accounts for 70% of each vector's squared length, the cosine similarity between two unrelated texts is approximately 0.7, and this ratio constitutes the model's noise floor.

If the cosine similarity between a large number of random, unrelated texts is consistently high (e.g., concentrated above 0.6), it indicates significant anisotropy in the embedding space.