Corey James- ~294,000structure-aware chunks from about 19,000 files in one code index
- 129configuration cells swept for the baseline's constants
- 5embedding models in the bake-off before the sweep
Two kinds of question, three streams
The memory service in FAFOHive serves agents that ask two kinds of question: an exact identifier such as a function name, and a description of behaviour in plain language. A dense vector search alone misses the first and lexical search alone misses the second, so I made code search and decision search hybrid.
A query runs up to three streams at once (dense similarity on embeddings from a multilingual embedding model, weighted full-text search, and a metadata stream) and the ranked lists are fused with reciprocal-rank fusion. A classifier looks at each query first and sets how much each stream counts, because identifier queries and descriptive queries need different blends.
exact-symbol handling lifts the chunk where it is defined
Holding up under agent load
After fusion, an optional cross-encoder rerank reorders the top results and a diversity pass drops near-duplicates, so an agent does not receive five copies of one paragraph. Capacity is part of the design: the embedding and rerank endpoints are bounded, and under load the reranker is the first stage dropped.
The graph side of the service is A code explainer that is not allowed to invent.
A baseline as strong as I could build
Capabilities model needed a bar to beat, and I ruled that the bar would be the strongest classical retriever I could build. That retriever is hybrid retrieval: dense embeddings fused with BM25 over the full tool catalog, with every weight and cutoff chosen by a sweep after an embedding-model bake-off.
I swept the baseline's constants, because a baseline tuned less carefully than the model it judges proves nothing.
Midway through the experiment a known-answer control failed to reproduce; the cause was a GPU precision setting that changed rankings relative to the CPU, and I pinned it off and re-ran the whole matrix. Every ranker in the experiment ran on both catalog sizes, so no comparison could pair a 42-tool model with a 2,000-entry baseline. The tie is the result I publish: The baseline that tied us.
Bars on a 0 to 1 scale.
The embedding models, the reranker and BM25 are third-party; the fusion, routing, capacity behaviour and measurement are mine.