Model Mechanics
Today
Explore
Practice
Library
Exam
Certificate
Search
⌘K
ខ្មែរ
◐
Sign in
Today
Explore
Practice
Library
Profile
Embeddings, Retrieval, and RAG
Today
5. Grounding, Adaptation, and Evaluation · 3 of 6
Mark read
Prev
Next
Orient
RAG is an evidence pipeline, not a truth switch
Find why the Khmer query retrieved an archived policy
Predict
Repair a bilingual policy retriever without leaking stale or private evidence
Close
Practice
Related
Embeddings, Retrieval, and RAG
Intermediate
☆
📑
Retrieve evidence first, then evaluate retrieval and generation separately
Orient
RAG is an evidence pipeline, not a truth switch
Find why the Khmer query retrieved an archived policy
Predict
Repair a bilingual policy retriever without leaking stale or private evidence
Close
Practice
Related
Practice activities and scenarios are optional. Mark the lesson complete after reading, and return to practice whenever it helps.
Restoring your saved work…
07
Optional practice
🧠 Optional scenario practice
Q1
The correct policy passage never appears in the top retrieved chunks. Which metric and component should be investigated first?
A
Final-answer preference scores and writing style, because a clearer generator can reconstruct evidence absent from context
B
Retrieval recall and the indexing/query/ranking pipeline
C
Embedding dimensionality alone, without testing chunking, metadata filters, query formation, index coverage, or ranking
D
End-to-end answer accuracy only, because separating retrieval from generation would make the product metric less realistic
Q2
Why can a RAG system still produce a false answer after retrieving the correct passage?
A
The passage may be longer than one paragraph, which prevents any language model from using it faithfully after retrieval
B
Dense retrieval optimizes only similarity, so the generator is required to choose the most popular claim regardless of evidence
C
Retrieval changes the model’s parameters for the current request, and the temporary weight update may encode the passage incorrectly
D
Generation may ignore, misread, or add claims beyond the evidence
08
Up Next
Advanced
Prompt, RAG, Fine-Tune, or Tool?
Beginner
Open-Weight and Local Deployment
Beginner
Model Lifecycle: Version, Monitor, Migrate
📝 Notes
Mark lesson complete
Next: Prompt, RAG, Fine-Tune, or Tool? →