Tuesday Noon: AI Vocabulary — Explaining What You Build
This morning you learned inference, fine-tune, embeddings, and hallucinate. Now we go deeper — the vocabulary you need when explaining your AI system to teammates, PMs, or stakeholders who aren’t deep in the weeds.
The hardest part isn’t building the RAG pipeline. It’s explaining it clearly.
🌟 Word of the Day: Grounding
| Pronunciation | /ˈɡraʊndɪŋ/ |
| Vietnamese | neo đậu / gắn kết vào dữ liệu thực — giúp AI trả lời dựa trên sự thật |
| Stress | GROUND-ing (2 syllables) |
🔊 Hear it: Cambridge · YouGlish
3 real-world examples:
- “We added grounding by connecting the model to our knowledge base — it no longer makes up answers.”
- “Without grounding, the chatbot was confidently wrong. With it, responses cite actual documentation.”
- “The key improvement in our agent was grounding every tool call result before the model synthesizes an answer.”
The simple explanation: Grounding = giving the AI something real to hold on to. Without it, the model floats free and invents things.
📋 Vocabulary: The RAG Pipeline in Plain English
| Phrase | Vietnamese | Example |
|---|---|---|
| RAG (Retrieval-Augmented Generation) | sinh text được hỗ trợ bằng truy xuất dữ liệu | ”We use RAG to make our chatbot answer from our actual docs, not from training data.” |
| chunk | đoạn văn bản nhỏ để embed và lưu vào vector DB | ”We chunk documents into 512-token pieces before embedding them.” |
| rerank | sắp xếp lại kết quả tìm kiếm theo relevance thực | ”After retrieval, we rerank the top 20 chunks to surface the most relevant 5.” |
| semantic search | tìm kiếm theo ý nghĩa, không chỉ từ khóa | ”Semantic search finds documents about ‘dog’ even if the query says ‘pet’.“ |
| context stuffing | nhồi nhiều thông tin vào context window | ”We’re context stuffing, which works but is expensive — we need better retrieval.” |
🗣️ Pronunciation Guide
Practice sentence (say this 3 times, slowly then faster):
“We use retrieval-augmented generation with semantic search and reranking to ground the model’s responses in verified documentation.”
Breakdown:
- retrieval = /rɪˈtriːvəl/ — re-TRIEV-al
- augmented = /ɔːɡˈmentɪd/ — aug-MENT-ed
- semantic = /sɪˈmæntɪk/ — se-MAN-tic
- reranking = /ˌriːˈræŋkɪŋ/ — re-RANK-ing
- verified = /ˈvɛrɪfaɪd/ — VER-i-fied
Tip: The key to sounding fluent with technical terms is rhythm, not perfection. Group related words: “retrieval-augmented generation” is one concept — say it as a unit, not three separate words.
✏️ Exercise 1: Fill in the Blank
Choose from: grounding, RAG, chunk, rerank, semantic search, context stuffing
- “We switched from keyword search to _______ so users can find answers even when they phrase things differently.”
- “The chatbot was hallucinating until we added _______ — now it always cites a real document.”
- “We _______ our PDF into 300-token segments before embedding each piece.”
- “Our retrieval step returns 50 results; then we _______ them and pass only the top 5 to the LLM.”
- “We’re _______ with the full codebase — 80k tokens per request. We need smarter retrieval.”
✅ Answers
- semantic search
- grounding
- chunk
- rerank
- context stuffing
✏️ Exercise 2: Translate to English
- “Chúng tôi dùng RAG để model trả lời dựa trên tài liệu thực tế, không phải dữ liệu training.”
- “Vấn đề là model đang hallucinate — chúng tôi cần grounding tốt hơn.”
- “Sau khi retrieve 20 chunks, chúng tôi rerank để chỉ pass 5 cái liên quan nhất vào context.”
✅ Suggested Answers
- “We use RAG so the model answers from real documentation, not training data.”
- “The problem is the model is hallucinating — we need better grounding.”
- “After retrieving 20 chunks, we rerank to pass only the 5 most relevant ones into the context.”
💡 Idiom of the Day: “Cutting through the noise”
Vietnamese: Lọc bỏ thứ không quan trọng để tìm ra điều cốt lõi
In AI conversations:
- “Reranking is our way of cutting through the noise — getting the signal from 50 search results down to 5.”
- “The hardest part of building a good chatbot isn’t the model — it’s cutting through the noise in the user’s question to understand what they actually need.”
💬 Mini Dialogue
Scenario: Explaining your RAG pipeline to a new backend engineer joining the team.
You: “Let me walk you through how our search works. When a user asks a question, we don’t just send it straight to the model.”
New engineer: “Oh — what do you do first?”
You: “We do semantic search across our knowledge base. The docs are chunked into 500-token pieces and embedded into a vector database.”
New engineer: “And then the model sees all the search results?”
You: “Not all of them — we rerank the top 20 and pass only the 5 most relevant into the context. This keeps costs down and improves grounding.”
New engineer: “Got it. So the model always answers from real docs?”
You: “Exactly. That’s the whole point of RAG — no hallucinations about things in our documentation.”
🎯 Challenge: Explain It in 3 Sentences
Right now, open a blank message draft and explain your current AI system (or any RAG system you know) in exactly 3 sentences using words from today:
- What the system does
- How grounding works in it
- What problem it solves
Don’t overthink it. Write fast, then read it aloud. That’s your practice for the day.
Tomorrow: Wednesday Noon — Architecture Vocabulary (system design terms for non-technical stakeholders) 🏗️