Vector Search
Vector search is a technique that converts data such as text, images, and audio into high-dimensional numerical vectors and retrieves semantically related content by computing inter-vector similarity metrics such as cosine similarity or Euclidean distance. Unlike traditional keyword matching, vector search captures semantic relevance, supports fuzzy queries and cross-modal retrieval, and is widely applied in information retrieval, recommendation systems, question-answering systems, and large-scale semantic search.
Core Facts
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.
向量数据库将文本内容通过嵌入模型（Embedding Model）转化为高维数值向量，检索时通过余弦相似度等算法找出语义最接近的内容
Vector retrieval captures semantic relevance rather than literal matching, which is its essential difference from keyword search.
Timeline (last 90 days)
The team uses the information extracted during the knowledge cleaning stage as metadata filter conditions for vector retrieval, so that recall is not only based on semantic similarity but can also be filtered by business attributes.
Vector retrieval (coarse ranking) is fast but lacks deep semantic understanding and may recall irrelevant content, so re-ranking is needed for fine ranking.
In practical vector retrieval systems, a globally fixed similarity threshold is usually set to filter out low-quality results, but the threshold makes it hard to adapt to the semantic needs of different queries.
When the knowledge base cannot answer a query, vector retrieval still returns semantically "seemingly relevant" fragments, so the large model may fabricate an answer as a result.
The system's knowledge base retrieval adopts a hybrid scheme of vector retrieval plus exact matching, first computing semantic similarity and then performing secondary ranking by urgency and knowledge freshness.
The limitations of vector retrieval include that retrieval quality depends on the semantic understanding capability of the embedding model, and it cannot guarantee the recall of all key information, posing a risk of memory omission.
The vector retrieval process is to split documents into chunks, use an embedding model to convert each chunk into a vector and store it in a vector database, and at query time convert the question into a vector as well and recall the closest several chunks by similarity.
Automatic linking in engineering usually relies on vector retrieval and semantic similarity computation: notes are split and converted into vectors, and connections are then established based on cosine similarity.
In scenarios with a very high proportion of unstructured text, the coverage of a knowledge graph may be less flexible than pure vector retrieval.
The example project WikiSearch's keyword index covers about 242,000 Simple English Wikipedia articles, while semantic search covers only about 25,000.
40 more timeline events
All Facts (20)
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.
80%Verified向量数据库将文本内容通过嵌入模型（Embedding Model）转化为高维数值向量，检索时通过余弦相似度等算法找出语义最接近的内容
80%VerifiedVector retrieval captures semantic relevance rather than literal matching, which is its essential difference from keyword search.
75%VerifiedSemantically similar texts are also close in distance within vector space, which is the mathematical foundation of vector retrieval.
70%VerifiedCursor uses an embedding model to transform the entire codebase into a searchable vector space to enable semantic search.
70%VerifiedIn RAG scenarios, if contradictory information exists in the index, vector retrieval cannot automatically determine which piece of content is more credible and will include both in the generation context, leading to confused answers.
65%VerifiedSemantic search can match results even when the query terms differ from the words in the file, as long as the meanings are similar, which is fundamentally different from keyword-based full-text search.
65%VerifiedVector retrieval based on semantic similarity is good at finding content that 'sounds similar' but struggles to capture causal chains, temporal dependencies, and hierarchical structures between pieces of information.
65%VerifiedKnowledge graphs are based on a graph-structured data structure, with construction complexity far exceeding vector-based solutions, but they can precisely express relationships between entities.
65%VerifiedImproving the distinctiveness of Skill descriptions can increase the hit rate; the writing approach should shift from 'what this Skill is' to 'when to use it.'
65%Verified关键词搜索基于精确字符匹配无法理解同义词，语义搜索依赖向量嵌入可捕捉语义关联，两者结合的混合检索效果优于单一方法
65%VerifiedIn actual production, hybrid retrieval combining both vector similarity and keyword matching is typically used.
65%UnverifiedAgent框架通过工具调用、检索增强生成（RAG）、记忆管理系统、规划与反思机制来弥补大语言模型的固有限制
90%Unverified向量检索无法精准匹配无语义的业务编号和专业术语，这是技术本身的能力边界而非实现缺陷
85%UnverifiedWhen memory scales to hundreds of MB or even GB, a pure file-based approach may still need to be combined with indexing mechanisms and vector retrieval technologies to balance simplicity and performance
60%UnverifiedModern RAG systems widely adopt vector retrieval technology, converting text into high-dimensional vectors via embedding models and computing cosine similarity to find relevant documents.
60%UnverifiedThe team uses the information extracted during the knowledge cleaning stage as metadata filter conditions for vector retrieval, so that recall is not only based on semantic similarity but can also be filtered by business attributes.
50%UnverifiedVector retrieval (coarse ranking) is fast but lacks deep semantic understanding and may recall irrelevant content, so re-ranking is needed for fine ranking.
50%UnverifiedIn practical vector retrieval systems, a globally fixed similarity threshold is usually set to filter out low-quality results, but the threshold makes it hard to adapt to the semantic needs of different queries.
50%UnverifiedWhen the knowledge base cannot answer a query, vector retrieval still returns semantically "seemingly relevant" fragments, so the large model may fabricate an answer as a result.
50%