The core finding
- A Google Research study cited in the video: when retrieval is off, the LLM doesn't give slightly wrong answers — it hallucinates more than if you had given it no context at all (
0:37). - The stated rule: bad retrieval is worse than no retrieval (
0:37).
Worked example — "What's your parental leave policy for employees in California?"
The system finds chunks mentioning parental leave and chunks mentioning California. Looks relevant. Three ways it silently fails (2:24):
| Failure | What happened |
|---|---|
| Stale version | The chunk comes from a 2019 policy updated twice since; the system only knows the words matched (2:46) |
| Bad split | The document was chunked mid-sentence, cutting off the eligibility criteria — the LLM gets half the information (2:46) |
| Destroyed structure | A table listing state-specific policies became a jumbled mess of words during text extraction (2:46) |
- In all three, search "worked" — it found text that looked relevant. The context was incomplete, outdated, or wrong (
3:10).
The counterintuitive part
- You'd expect the LLM to say "I don't know" when the context is insufficient. It doesn't (
3:10). - Giving an LLM context makes it more confident — even when the context is bad. It fills gaps with information that sounds right but is completely made up (
3:26). - Result: a confidently wrong system, far worse than one that admits it doesn't know (
3:36).