Embedding Model

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

Vector memory databases convert historical conversation content into high-dimensional semantic vectors, and during new conversations dynamically retrieve relevant memory fragments via similarity search and inject them into the context window.

嵌入模型将文本转换为高维向量时，通常生成768维或1536维的浮点数数组

Vector databases (such as Pinecone, Weaviate, Milvus, etc.) convert text into high-dimensional vectors via embedding models and use cosine similarity or Euclidean distance for approximate nearest neighbor (ANN) search.

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

RAG tutorials rarely cover a key issue: chunk boundaries are almost always wrong; when documents are chunked and embedded, assumptions are actually being made about which context belongs together, and these assumptions often do not hold.

If a RAG pipeline returns results that are technically correct but contextually useless, the problem is almost certainly in the chunking or metadata, not the embedding model.

The semantic caching mechanism vectorizes and stores historical responses to similar requests, and when semantic similarity exceeds a threshold, it directly returns cached results to reduce upstream API call frequency and costs

Vector embedding is the process of transforming discrete text, code, or experiences into numerical vectors in a continuous high-dimensional space, so that semantically similar content is closer in vector space.

Long-term memory retrieval relies on vector database technology at its core, converting text into high-dimensional vectors via embedding models.

The limitation of specialized small models is that their generalization ability rapidly degrades when task boundaries are ambiguous, cross-domain reasoning is required, or unseen edge cases need to be handled.

In code scenarios, slicing by function boundaries preserves more complete semantic units than slicing by fixed character counts.

The RAG strategy encodes code snippets as high-dimensional semantic vectors through vector embeddings, and during retrieval computes the semantic similarity between query intent and code snippets to recall relevant code.

OpenAI's text-embedding-ada-002 can map text into a high-dimensional vector space, where semantically similar texts are closer together in the vector space.

Embedding models typically compress text into dense vectors of 768 or 1536 dimensions.

Embedding models are neural network models that map discrete inputs into dense vectors, with representative examples including OpenAI's text-embedding-ada-002, Voyage AI's voyage series, and open-source BGE, E5, GTE and others.

The experiment used open-source or low-cost embedding models to convert input text into vectors, then fitted a simple sklearn model on top of the embeddings and compared it with Jev in terms of effectiveness and cost.

The embedding + sklearn solution is cheaper than Jev and almost always significantly more accurate.

Embedding models are only responsible for representation, have fewer parameters, low inference cost, and do not depend on prompt engineering.

When labeled data is available, embedding + sklearn is usually better, especially when the amount of labeled data is sufficient to cover the task distribution.