Web architecture

LLM in Software engineer’s view

• LLM is a stateless(no memory between request) HTTP API endpoint，接收prompt并返回text

Hallucination Problem

• LLMs generate plausible-sounding but factually incorrect response

RAG

• Retrieve
• Augment
• Generate

semantic search vs. traditional datbase queries

FIne tuning vs. API+RAG

Chunking

Chunking required

Recursive character text splitting

Token Overlap is Mathematically Required

Metadata

Bulk asnyc parsing pipeline

• 批量异步解析pipeline

Pipeline Error handling

summary

Embedding

definition

• embedding是固定长度的浮点数array表示text的semantic meaning

Cosine Similarity

Hierarchical Navigable Small World

Vector Search

Lanchain.js

---

链接