Black Box AI

Aviation safety questions split into two very different types. Counting questions need exact structured data, while causal or explanatory questions need retrieval over accident narratives.

A single chat-style interface can hide that distinction and make answers feel more certain than they are. The project needed to route questions to the right engine, show the evidence, and keep generated prose from inventing numbers.

Black Box AI routes plain-English questions to SQL, retrieval, charting, or a combined path. Structured questions use deterministic plans or guarded model-planned SQL, then return the SQL, table, chart, and audit trail. Narrative questions retrieve cited accident reports and show the source records directly.

The live app includes FastAPI, React, provider adapters, in-memory BYOK handling, query planning, SQL validation, chart validation, retrieval, answer composition, limitations, and an audit trail for each answer.

The retrieval layer was scored at the accident level across semantic, BM25 keyword, and reciprocal-rank-fusion hybrid retrieval, on 17 hand-labeled questions. The result contradicted the starting hypothesis, which is the reason it is reported.

Hybrid lost on recall. Reciprocal rank fusion is a consensus mechanism, not a best-of-both. It helps when two engines are each partially right and complementary, and it actively hurts when one is decisively right and the other decisively wrong, because the confident wrong results push the correct ones down. On the bird-strike query semantic scored 0.80 and keyword 0.40; hybrid returned 0.20. Fusion had little to resolve here because on this corpus the search term usually appears in the report text, so the two engines mostly already agreed.

Two of the three metrics carried no signal, and Hit@5 was the worst offender. It saturated at 1.000 for every engine, because the corpus is dense enough that finding something relevant in the top five is trivial. Recall@5 was also capped: three topic-saturated queries (hard landing, carburetor icing, gear collapse) each have twenty-plus relevant reports, which holds Recall@5 near 0.25 for any retriever. MRR is the only metric here that discriminates, and it separates the engines by 4.9 MRR points across a 17-question set, which is not enough to call decisively.

Conclusion, and the design decision that followed. Hybrid retrieval is insurance against query types you cannot anticipate, not a universal upgrade. On a corpus where the answer term reliably appears in the text, a single retriever is competitive, so the demo defaults to semantic and exposes all three for comparison. That was a decision driven by measurement rather than by assumption, and the measurement said the opposite of what I expected.

What would strengthen it. Seventeen queries cannot separate methods whose real differences are a few percentage points, and the set is too easy: it was written by someone who knew the corpus. A stronger set would target each engine's known failure mode deliberately, with paraphrase queries that share no vocabulary with the report, rare-identifier queries such as tail numbers and exact dates, and multi-hop questions requiring two documents. It would also report bootstrap intervals rather than point estimates. That is the next piece of work on this project.