RAG Chatbot: Building a PM Interview Knowledge Engine

From Static Prompts to Retrieval-Augmented Generation

My portfolio already had an AI chatbot — a prompt-engineered assistant that could talk about my work. But I wanted something sharper: a second mode that answers PM interview questions grounded in my actual experience, not general knowledge. This is how I built a full RAG pipeline using Cloudflare's AI stack.

---

The Problem: Generic Answers to Specific Questions

The original chatbot was effective for portfolio conversations — "Tell me about your projects" or "What's your tech stack?" — but it struggled with deeper PM interview questions. Ask it "How did you handle a scaling challenge?" and it would give a reasonable but generic answer, drawing loosely from its baked-in knowledge base.

I needed answers that cited specific metrics, timelines, and outcomes from my real experiences. Not "I believe in data-driven decision making" but "Our settlement table hit 45 crore rows, so I audited 30+ queries and implemented date-based partitioning with a 10-minute downtime window."

---

The Solution: A Dual-Mode Architecture

Rather than replacing the existing chatbot, I built a toggleable dual-mode system:

• Chat Mode — The original prompt-based assistant. Warm, conversational, great for general portfolio questions.
• PM Expert Mode — A RAG-powered engine that retrieves relevant experiences from a vector database and answers only from those. Factual, structured, interview-ready.

The difference in answers is stark — by design.

---

How It Works Under the Hood

The RAG pipeline runs entirely on Cloudflare's edge infrastructure:

1. Knowledge Base → Vectors: I distilled 25 PM experience "atoms" — each a self-contained story (scaling challenges, stakeholder conflicts, process redesigns, vendor negotiations). These are embedded using Workers AI (bge-base-en-v1.5, 768 dimensions) and stored in Cloudflare Vectorize.

2. Query → Retrieval: When a user asks a question in PM Expert mode, their query is embedded in real-time and matched against the vector index. The top 5 most relevant experiences are retrieved, filtered by a similarity threshold to avoid noise.

3. Context → Generation: Retrieved experiences are injected into a carefully engineered system prompt, and Groq's Llama 3.3 70B generates the response — grounded exclusively in the retrieved context.

4. Full Observability: Both modes have complete Langfuse tracing — embedding latency, vector search scores, LLM generation metrics — so I can compare performance across thousands of conversations.

---

The PM Thinking Behind It

Several decisions here were driven by product instinct, not just engineering:

• Preserving Both Modes: Users visiting my portfolio want different things. A recruiter browsing casually benefits from Chat mode. Someone preparing for a PM interview discussion benefits from PM Expert. Forcing one mode would underserve half the audience.
• Prompt Engineering for Tone: The hardest part wasn't the RAG pipeline — it was making the LLM sound like me in an interview, not a document retrieval system. The prompt explicitly bans generic conclusions like "This taught me the importance of..." and instructs the model to lead with the most impressive metric first.
• Edge-First Architecture: The entire pipeline — embedding, vector search, and routing — runs on Cloudflare's edge. No cold starts, no external vector DB, no separate embedding service. Latency stays under 2 seconds.

---

Try It Yourself

Open the chat widget on any page of this website. Toggle to PM Expert mode and try:

• "How did you handle a database scaling challenge?"
• "Tell me about a stakeholder conflict you resolved"
• "Are you a good PM?"

Then switch to Chat mode and ask the same questions. The difference is the point.

---

Technical Stack

• Cloudflare Vectorize — Vector storage and similarity search
• Cloudflare Workers AI — Embedding generation (bge-base-en-v1.5)
• Groq + Llama 3.3 70B — LLM inference
• Langfuse — Observability and tracing
• Cloudflare Pages Functions — API proxy layer