Mode
Parameters
How documents are split for embedding
Number of retrieved chunks
Min cosine similarity for inclusion
Presets
Pipeline Flow
Step 0 / 0
Knowledge Base Chunks
What's Happening
Click Next Step to begin stepping through the RAG pipeline. Each stage shows exactly what data flows and why it matters.
| Aspect | No RAG | Basic RAG | Agentic RAG |
|---|---|---|---|
| Knowledge | Static pretrained data only | Retrieved docs augment context | Iteratively refined retrieval across sources |
| Query Handling | Direct to LLM, no preprocessing | Single embed + retrieve pass | Query refinement + multi-pass search |
| Hallucination Risk | High | Medium | Low |
| Citations | None | Possible with source metadata | Built-in with provenance tracking |
| Response Quality | Generic / may be wrong | Grounded in retrieved facts | Synthesized from multiple sources |
| Best For | General chat, creative tasks | Factual Q&A, customer support | Complex research, multi-step reasoning |
1. No RAG at All
The book's PoC failure story: "A successful PoC that works on clean data in a lab is not a guarantee of a production-ready system." The agent hallucinated a return policy that didn't exist.
2. Poor Chunking
"The quality of your retrieval is paramount. Invest in a robust document processing pipeline that cleans and chunks your source data effectively." Poorly chunked docs lose context and break retrieval.
3. Ignoring "No Match"
When retrieval returns nothing relevant, the agent should say "I don't know" instead of hallucinating. Book example: agent confidently told a customer they were eligible for a refund that didn't exist.
4. Static Knowledge Only
RAG's power is access to CURRENT information. The book contrasts "Static, pretrained model data" vs "Dynamic, external knowledge base (service bulletins)."
5. Not Citing Sources
"This often allows the agent to provide citations or references back to the source documents, which dramatically increases transparency and user trust." Always surface where retrieved info came from.
Step through the pipeline to generate the augmented prompt...