How Ask Amalachukwu Works
This is the model card for the portfolio's own AI assistant -- meant to demonstrate the same transparency this site advocates for elsewhere.
- Purpose
- Answer recruiter/visitor questions about Amalachukwu's professional evidence, citing sources or saying plainly when no evidence exists.
- Knowledge sources
- Exclusively this site's structured evidence data (lib/evidence). Currently 16 verified record(s) and 0 pending-verification placeholder(s) that carry no content.
- Retrieval boundary
- No arbitrary web access, no filesystem access, no shell or tool execution, no database beyond the static evidence content. The assistant cannot answer from anything outside that list.
- Provider
- Currently a deterministic mock provider (no real model call) -- see lib/ai/provider.ts. A real provider will be wired in behind the same interface once there is a verified corpus worth answering from.
- Hallucination controls
- The system prompt instructs the model to answer only from the supplied evidence list and to say so when it can't. This is a prompt-level control, not a guarantee -- see Limitations.
- Privacy & logging
- Questions are not persisted. Only a per-IP request count is kept, in memory, for rate limiting.
- Rate limiting & abuse controls
- 10 requests per minute per IP address (in-memory; see /security for the noted durability limitation), 2,000-character input limit, 15-second request timeout.
- Limitations
- No real model is connected yet, so no response currently demonstrates actual language understanding. Once a real provider is wired up, prompt-level instructions to stay grounded are not a hard guarantee against hallucination -- this page will be updated with whatever additional controls (e.g. citation verification against the evidence list before returning a response) are added at that point.
- Fallback behavior
- An unrecognized provider configuration falls back to the mock provider rather than failing the request outright.