Resume RAG: Ask my resume

An /ask page on the resume site that lets visitors query my experience, projects, and skills in natural language. The whole RAG pipeline runs on Cloudflare, and is itself a demonstration of building LLM / RAG applications:

• Workers AI · bge-m3: multilingual embeddings (1024-dim) for the question and resume content
• Vectorize: vector search that retrieves the most relevant resume chunks (top-k)
• Workers AI · Llama 3.3: streams an answer grounded in the retrieved chunks
• Astro Pages: bilingual chat UI with token-by-token streaming and cited sources
• Deploy-time ingest: resume content is chunked → embedded → upserted into Vectorize

The focus is trustworthiness and cost control: the safeguards a public, unauthenticated endpoint needs:

• Anti-hallucination grounding: answers only from retrieved chunks; says “not in my resume” when it can’t find something instead of guessing; declines sensitive questions (salary / personal data)
• Prompt-injection defense: the user’s question is wrapped in <question> fences so it can’t override the rules
• Cost guardrails: a global daily cap plus per-IP rate limiting, so it can’t be abused as a free LLM proxy
• Graceful degradation: AI / vector failures return a friendly error rather than a bare 500