Retrieval training gets interesting once the model is strong enough to create its own mistakes. The usual static negative set is too easy; the index has already moved on.
The loop I want for HOBIT-style retrieval is:
- Encode the corpus with the current retriever.
- Retrieve near misses for each query.
- Mix labeled positives with hard negatives from the live index.
- Re-train with a loss that keeps the margin honest.
- Rebuild the index and repeat.
Why Hard Negatives Matter
Easy negatives make the loss look healthy while recall barely moves. The useful examples are the documents that are close enough to fool the current model but wrong enough to teach it a boundary.
Retrieval mixer
HOBIT training loopHard negatives are drawn from the current retriever so the model learns the boundary it actually uses.
Hybrid lexical and dense searchA lexical retriever finds exact anchors while a dense retriever recovers paraphrases and concept matches.
PageRank at web scalePower iteration turns link structure into a stable authority score under a random surfer model.
PPO entropy collapsePolicy updates can become too certain when advantage estimates and clipping remove exploration pressure.
Measuring the Frontier
The model is only part of the system. Index latency, candidate count, and reranker depth all move the frontier together.
Recall and latency frontier
Notes to Expand
- Add a real experiment table for recall, NDCG, and p95 latency.
- Separate mined negatives from synthetic query negatives.
- Track which failure clusters survive multiple training rounds.