Ask Zico
A launched source-grounded Arabic assistant built around controlled retrieval, conservative answers, and evidence-first UX.
The work
Ask Zico is the Arabic-first assistant I built for Madraset El Shamamsa. It started in May 2026 and launched publicly on July 2026.
The product answers from the site's public educational material, returns the sources that support an answer, and falls back to source links or search when the evidence or model capacity is not strong enough. The central constraint is deliberate: model fluency never outranks evidence identity.
What I built
-
Developed a retrieval-augmented generation assistant with semantic retrieval, domain-aware routing, lexical fallback, and sources tied to retrieved evidence. The server accepts only citation IDs that resolve to retrieved trusted content.
-
Built the edge backend on Cloudflare Workers, using Vectorize for semantic candidates, KV for lexical and source artifacts, and D1 for quota reservations, feedback, and operational events.
-
Improved reliability with a quota-aware Gemini model ladder, controlled OpenRouter fallback for provider failures or exhausted capacity, structured answer validation, rate limits, and source or search fallback when generation is unavailable or unsupported.
-
Implemented an authenticated PHP-to-Worker boundary, privacy-aware observability, and evaluation tooling that supports regression testing and deliberate retrieval iteration.
Read the series
The articles below document dated evaluation checkpoints, implementation decisions, and failure cases rather than presenting historical measurements as permanent product guarantees. Sensitive production traffic, prompts, identities, limits, and credentials are intentionally excluded.
Start with the production architecture.
Architecture and decisions
RAG quality
- Citation-ready ingestion
- Controlled hybrid Arabic retrieval
- Grounded answers and guardrails
- Retrieval, answer, citation, and refusal evaluation