RAG & groundedness
P11.ir-rag.04 · Audience: guest, it-ml, language-pro · Prerequisites: Information retrieval & hybrid search, Re-ranking & IR metrics
A plain LLM answers from frozen, unattributable parametric memory — and is
confidently wrong when the answer isn't in it. Retrieval-augmented generation
grounds the answer in retrieved passages: fresh, cited, and able to abstain.
Most RAG failures are retrieval failures — which is why P11.ir-rag.01/.02 carry
the load. (Cross-link: P10.llm-essentials.03 — reliability & grounding.)
RAG fixes three failures of a plain LLM at once:
- Fresh — retrieve from a live index, so the answer reflects today's data, not the training cut-off.
- Attributable — the answer cites the retrieved passages it used.
- Grounded — the model answers only from those passages, and abstains when they don't support an answer.
question ─▶ retrieve (IR1) ─▶ re-rank (IR2) ─▶ [context] + prompt ─▶ LLM ─▶ grounded answer + citations └▶ abstain if context is weak
🎯 RAG = grounding an LLM in retrieved content. The generator is the easy part; the retriever, the grounding rule, and the abstention rule make the answer trustworthy.
Ask the mentor about this module
Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.
🎓 Practice ladder
3 graded rungs · ~33 minNow build the RAG pieces yourself. Each rung is a three-panel workspace: instructions on the left, a code editor in the middle, output + test results on the right. Run checks the visible tests; Submit grades against hidden edge cases — overlap ≥ chunk size must raise, the out-of-scope query abstains, a threshold sweep flips answer↔abstain, and faithfulness averages only the answered cases. The senior rung rebuilds the whole groundedness evaluator.
Rung 1 — Chunk with overlap
Loading exercise…
Rung 2 — Answer or abstain
Loading exercise…
Senior rung — Score the RAG pipeline
Loading exercise…
⚡ Interview Ref — the quick-scan Reference face
- RAG = grounding an LLM in retrieved content: fresh (live index), attributable (citations), grounded (answer only from context, else abstain).
- Chunking: size trade-off (dilution vs lost context) + small overlap for straddling facts; metadata for pre-filters + citations.
- Pipeline: chunk/filter → hybrid retrieve → re-rank → generate. Most RAG failures are retrieval failures.
- Grounding rules: constrain to context · cite · abstain below threshold.
- Eval: faithfulness · answer relevance · retrieval recall (the ceiling) · abstain rate (healthy > 0). Recall bounds the rest — fix retrieval first.
- Healthcare: hard grounding, conservative abstain+escalate, audit trail, PII/PHI at every hop.
Try it yourself
A scratch console for this page's ideas — ungraded, nothing you run here is recorded.
Scratch console
A scratch console with the scientific stack (pandas, numpy, scikit-learn). Runs on the server — no network, resource-limited and measured.
Output appears here.