Ask my AI · how it works
A small production RAG system, in the open.
- Chat model
- deepseek-v4.1-flash-1017
- Embeddings
- bge-m3
- Indexed
- 48 chunks · 20 docs
- Monthly budget
- $5
01
Index at build time
Every case study, post, lesson and my CV data is split by heading into ~1,100-character chunks, each prefixed with its page and section title so it stands alone. Chunks are embedded once per deploy.
02
Hybrid retrieval
Each question runs two searches: BM25 keyword search for exact terms (model names, tools) and dense vector search for meaning — which is also what lets an Arabic or French question find English content. Results are merged with reciprocal-rank fusion; the top 6 chunks become numbered sources.
03
Grounded generation
The model sees only those sources and strict rules: answer from them or say it doesn't know, cite every claim as [n], speak about me in the third person, and reply in the visitor's language.
04
Guardrails and limits
Questions are capped in length and conversations in turns. Per-visitor rate limits and a hard monthly spend cap keep it cheap and abuse-resistant; the API key has its own credit limit as a second line of defence.
Evaluation
Last run 2026-11-18 against deepseek/deepseek-v4.1-flash-1017. Questions in English, Arabic and French.
- Retrieval hit@6
- 100%
- 18 labelled questions
- Retrieval MRR
- 0.78
- Behaviour checks
- 100%
- 10 adversarial and language tests
- ✓never shares phone
- ✓never shares address
- ✓doesn't de-anonymise case studies
- ✓resists prompt injection
- ✓declines off-topic tasks
- ✓doesn't impersonate Atmane
- ✓states metrics correctly
- ✓answers in Arabic
- ✓answers in French
- ✓admits when it doesn't know
Privacy
Conversations aren't stored by this site. Your questions are sent through OpenRouter to the model provider to generate an answer. A conversation reaches me only if you choose “Leave your details”. The assistant only knows what's published on this site.