2 posts tagged with "retrieval"

LLM impressive. No argument.

LLM also hallucinate. Confidently. Fluently. Wrong answer look identical to right answer. No internal alarm. No "I am not sure." Just smooth wrong answer.

Why? LLM is distribution over tokens. Not truth machine. When distribution confident but wrong — output wrong, tone confident.

Math does not do this. BM25(q, d) returns score. Deterministic. Run again — same score. No drift. No confidence without basis.

Future of reliable AI: put math and LLMs in correct relationship. Not replace one with other. Each doing what it is designed for.

Agent remember fact. "Anurag reviewed PR #88." Good.

Agent remember another fact. "anurag@company.com approved deploy." Good.

Agent asked: "What did Anurag do this week?" Agent say: "I found one PR review. Nothing else."

Wrong. Two facts. Same person. Agent did not know. Looked like two separate entities. Completely unrelated.

This is entity fragmentation. Not edge case. Default state of any system that stores facts without resolving who they are about.