29 core terms
AI support glossary
Core terms in the open-source AI support space — RAG, Embedding, MCP, Reranker, Faithfulness and more.
Core concepts 10
- RAG Retrieval-Augmented Generation
- Retrieval-Augmented Generation. Recall relevant chunks from a knowledge base via vector or keyword search, then have the LLM answer using those chunks. Nearly all AI support stacks use it.
- Embedding 词向量 / 文本嵌入
- Mapping text into a high-dimensional vector space where semantically similar texts are close. For Chinese, bge-m3 and Conan-embedding are common.
- Vector DB 向量数据库
- A database optimized to store vectors and search by similarity. Common: Weaviate, Milvus, Qdrant, LanceDB, pgvector.
- Reranker 重排序模型
- Model that re-scores initial retrieval results. Significantly boosts top-1 accuracy. For Chinese, bge-reranker and bce-reranker are common.
- MCP Model Context Protocol
- A protocol from Anthropic (2024) standardizing how LLMs interact with external tools and data sources. Supported by Dify, Open WebUI, LibreChat.
- Function Calling 函数调用
- The LLM's ability to decide which external function to call based on user input and return structured arguments.
- Agent 智能体
- An LLM application pattern where the model plans, calls tools, and reflects on results autonomously — versus simple Q&A. Dify, LangGraph, AutoGPT are Agent frameworks.
- Workflow 工作流
- A graph-style orchestration of multi-step logic (branches, loops, HTTP calls). Dify, FastGPT and n8n all ship workflow features.
- LLM Large Language Model
- Large Language Model. Common LLMs discussed here: GPT, Claude, Qwen, DeepSeek, GLM, ERNIE, Doubao.
- Prompt 提示词
- Instruction text sent to an LLM. Support prompts typically include role, style, constraints, and KB context.
RAG 3
- Chunking 分块
- Splitting long documents into LLM-friendly chunks. Strategies include character count, paragraph, heading, QA-split, parent-child.
- GraphRAG
- A RAG variant that extracts knowledge into a graph before retrieval. Strong on entity-rich domains like finance and healthcare. Built into RAGFlow.
Evaluation 4
- MRR Mean Reciprocal Rank
- Mean Reciprocal Rank. Evaluates retrieval — the earlier the first correct answer appears, the higher the score.
- Faithfulness 忠实度
- Whether the answer is grounded entirely in retrieved content (no hallucination). The key generation-side metric.
- Hallucination 幻觉
- LLM-generated content that is factually wrong or unsupported. Support AI must be tightly constrained to reduce hallucinations.
Operational KPIs 5
- Deflection Rate 自助解决率
- Share of conversations resolved by AI without human handoff. A core AI-support KPI, typically targeted 45-70%.
- CSAT Customer Satisfaction
- Customer satisfaction rating (1-5). Collected post-conversation; track AI leg and human leg separately.
- AHT Average Handling Time
- Average handling time per conversation. Drops noticeably after AI is introduced.
- FRT First Response Time
- First Response Time. The interval from a user's first message to the first reply. AI support can bring it to seconds.
- SLA Service Level Agreement
- Service Level Agreement — committed FRT, resolution times etc. Chatwoot supports tiered SLAs per customer.
Platform features 4
- Inbox
- Chatwoot's channel abstraction. Every channel (web, email, WhatsApp) is an Inbox with team / auto-assignment / SLA config.
- Agent Bot
- Chatwoot's external AI hook. Sends messages to an external service (like Dify) via webhook for replies.
- Pipelines
- Open WebUI's Python middleware. Lets you inject business logic before / after the LLM call.
Infrastructure 3
- Ollama
- Local LLM inference engine — the most popular open-source OpenAI-API drop-in. Mac / Linux / Windows.
- vLLM
- High-throughput LLM inference framework. Production-grade; much higher throughput than Ollama.
- OpenAI-compatible API
- API compatible with OpenAI's schema. Many open-source models / providers offer it (DeepSeek, SiliconFlow, Ollama) — trivial to swap.