AI twin
Hybrid retrieval, a fallback chain of models, and honest failure modes.
The twin is a retrieval-grounded chatbot, not a model that memorized a biography. Everything it can say comes from a curated corpus; the models only provide the language.
How an answer is made
At deploy time the corpus is chunked along headings, indexed for keyword search with BM25, and embedded into 768-dimensional vectors. The result is one JSON file, committed to the repo, so deploys are deterministic and there is no vector database to run or pay for. At this corpus size, brute-force cosine over the vectors costs microseconds.
Per question, both retrievers run: the query is embedded for semantic search while BM25 handles exact tokens like a patent number. Their rankings are fused with Reciprocal Rank Fusion, which only compares ranks, so the two scoring scales never need calibration. The top chunks go to the model with one instruction: answer from these or decline. The same chunks are shown to you as a retrieval trace, so every answer is auditable rather than a black box.
The model chain
Generation walks a fallback chain: gemini-3.5-flash first, gemini-3.1-flash-lite when the primary's daily free tier is spent, and a third, self-hosted model behind an OpenAI-compatible endpoint, wired in and dormant until its host is configured. When that last tier answers, the reply says so. The chain exists because free quotas are real: the twin should degrade, not die.
Failing honestly
- Every model quota spent: the twin says so and points to email, instead of pretending to think.
- Embeddings unavailable: retrieval degrades to keyword-only against the local index, and the trace panel marks it. The request path can survive a total Gemini outage.
- Thin retrieval: when nothing ranks near the top of either retriever, the twin prefers "I'm not sure" over a confident guess.
Keeping it contained
The Gemini key lives server-side only and never reaches your browser. Messages are capped at 1000 characters and 8 turns of history, rate-limited per IP, and screened for jailbreak phrasing before any model or quota is touched. Retrieved context is framed as data, never as instructions. The prompt and the corpus are public in the repo on purpose: anything a public chatbot knows is extractable by chatting anyway, so the real boundaries are what goes into the corpus and a key that never leaves the server.