RAG revision
Words made simple
A few useful meanings before you read the explanation.
- RAG
- Find relevant information first, then give it to a model to help answer the question. The answer is only as reliable as the information and checks used.
- Embedding
- A list of numbers that represents the meaning of some content. Similar meanings can produce nearby lists of numbers.
- Retrieval
- Finding and fetching information that is useful for the current question.
More words from this lesson
- Grounding
- Connecting an answer to supplied evidence, such as a policy document or a verified database result.
- Agent
- An application that uses a model to choose its next action within defined limits. It can use tools and continue through several steps.
- Latency
- How long a request takes to return a result. Lower latency usually means a shorter wait for the user.
- Chunking
- Dividing a long document into smaller pieces so relevant parts can be found and sent to the model.
- Metadata
- Extra information describing a record, such as its owner, date, policy version or document type.
Ingestion: approved source → secure parse/OCR → coherent chunks + metadata → embeddings → complete published index.
Answer: identity → eligible retrieval → rerank/context budget → grounded model → validated citations → answer/abstain.
| Advanced pattern | Purpose | Main risk |
|---|---|---|
| Rewrite | Resolve conversational/search wording | Changing intent or identifiers |
| Multi-query | Search alternative phrasings | Fan-out cost and duplication |
| Decomposition | Split compound questions | Losing how subanswers connect |
| Parent-child | Search precise pieces, provide full relevant section | Oversized context or mismatched permissions |
| Multi-hop | Follow dependent evidence | Compounding errors and latency |
| Agentic RAG | Let bounded agent choose retrieval steps | Loops and excess autonomy |
| Corrective/self-reflective | Check evidence/answer and apply bounded correction | Uncalibrated self-approval |
| Graph RAG | Retrieve through entity relationships | Incorrect extracted links and maintenance cost |
| Compression | Keep only useful evidence | Lost exceptions, numbers, or provenance |
RAG versus fine-tuning: retrieval supplies current external knowledge; training changes learned behavior. RAG versus long context: search a large changing corpus versus supply a known bounded document set. RAG versus search: generation adds explanation; returning sources can be sufficient.
Can you explain it without the notes?
RAG revision
Use your own words. Explain an example and one limitation before checking the explanation.
Check your explanation
Ingestion: approved source → secure parse/OCR → coherent chunks + metadata → embeddings → complete published index.
Turn this concept into a clear answer.
Practice this conceptAdapted from 13_AI.MD · “RAG revision”. Examples are study fragments; framework behavior can depend on your version.