Vector Embedding

AI model performance is positively correlated with the amount of data fed to it, and enterprises must continuously input business details to obtain more accurate AI outputs

RAG（检索增强生成）技术架构将文档切片、向量化、存入向量数据库，查询时先检索相关片段再交给模型生成回答

The working principle of RAG is divided into a retrieval phase (converting the question into a vector to recall relevant content from the knowledge base) and a generation phase (inputting the recalled context together with the question into the model to produce an answer).

The core of semantic search is vector embedding technology, which converts code snippets and natural language queries into high-dimensional vectors and calculates semantic distance using methods such as cosine similarity.

大语言模型的预训练阶段本质上属于自监督学习，是一种特殊的非监督学习，模型通过预测下一个词来从海量文本中自动学习语言规律

BM25 and TF-IDF are classic representatives of lexical retrieval, based on literal keyword matching and term frequency statistics.

Vector embeddings typically map text into high-dimensional spaces of 512 to 4096 dimensions.

AI应用中的向量通常是几百维到几千维的高维数据，维度越高承载的语义信息越丰富

A Transformer maps tokens to high-dimensional vectors, and semantic relationships manifest as geometric structure in the vector space—similar concepts are closer together, forming clusters and manifolds.

Semantic similarity retrieval measures topic relevance rather than answer entailment; a passage may be topically relevant yet lack the specific facts needed to answer the question, which is the direct cause of inducing unsupported inferences in language models.

Vector embeddings, AST parsing, and knowledge graphs are complementary: vector embeddings are responsible for "finding it," AST parsing for "understanding the structure," and knowledge graphs for "tracing the relationships."

Repository intelligence is a collection of technologies, including vector embeddings, AST (abstract syntax tree) parsing, and knowledge graphs, whose purpose is to let AI coding agents understand the structure, dependencies, and relationships of an entire codebase, not just the file currently open.

Vector embeddings convert code blocks into numerical representations and retrieve the parts most semantically similar to the agent's current task, essentially moving retrieval-augmented generation (RAG) from documents to source code.

Vector embeddings measure "how similar" rather than "whether connected," and therefore provide no guarantee for determining calls and dependencies.

Vector embeddings map text fragments to a set of floating-point numbers in a high-dimensional space, where semantically similar content is closer together in that space.

For small-to-medium-scale scenarios, directly pasting relevant text into the prompt is often simpler and no less effective than building a RAG system based on vector embeddings.

Some cross-platform AI note-taking tools support semantic search, meaning users can query historical meeting content directly using natural language questions, a capability that can only be fully leveraged when data is centralized.

LangGraph's persistent memory is typically stored in the form of vector embeddings, and retrieval relies on cosine similarity or approximate nearest neighbor (ANN) algorithms to match semantically relevant memory entries.

File Reasoning capability relies on technologies such as OCR, layout analysis, table extraction, vector embedding, and semantic search to achieve understanding and analysis of unstructured files.

Vector embedding retrieval faces a "semantic gap" problem when dealing with enterprise-level multi-repository and complex call relationships — it essentially "guesses" the associations between code rather than precisely knowing the call and dependency relationships.