PitchPilot — a field-sales companion with a RAG copilot (case study)

Source is private. This is the architecture and the thinking behind it — no product code. The retrieval technique is shown, runnable, in a separate repo: rag-demo.

PitchPilot is a white-label app for door-to-door / field sales reps: capture a lead, run the pitch, handle objections, and track commission — all on a phone, at the door. Its core is Wingman, a retrieval-augmented copilot that answers a rep's question ("what do I say when they mention price?") from the product's own knowledge base, in real time, grounded and cited.

The problem

A field rep can't read a 40-page playbook on a doorstep. When a prospect pushes back, the rep has seconds to answer, accurately, in the product's own words. Wingman turns the whole playbook — pricing, objections, scripts — into something you ask in natural language and get a one-paragraph, on-message answer from.

Architecture

flowchart LR
    subgraph Ingest [once, offline]
        D[product docs<br/>pricing · objections · scripts] -->|chunk + embed| KB[(pgvector KB)]
    end

    subgraph Ask [at pitch time]
        R[rep question] --> EF[copilot-ask<br/>edge function]
        EF -->|embed + match| KB
        KB -->|top-k chunks| EF
        EF -->|context-limited prompt| G[Gemini]
        G -->|grounded answer + actions| R
    end

Loading

• Knowledge base — the playbook is chunked, embedded, and stored in Supabase pgvector, partitioned by product so a white-label tenant only ever retrieves its own content.
• Retrieval RPC — a Postgres match function returns the nearest chunks; this is the exact shape shown in rag-demo.
• Grounded generation — a Supabase Edge Function builds a context-limited prompt and calls Gemini. The system prompt hard-limits the model to the retrieved chunks and forbids inventing pricing or promises — accuracy over hype, because a wrong answer at a doorstep costs a sale.
• White-label — one codebase rebrands for any industry in under a day; a public demo runs as "VoltLine" for the energy sector.

What made it actually work (the debugging that mattered)

• Retrieval was silently returning nothing. An IVFFlat index over a tiny corpus collapsed recall to zero — the copilot kept deflecting to "check the FAQ." Dropping to exact search fixed it instantly. Lesson: approximate indexes need enough rows to be approximate over.
• Truncated mid-sentence answers. The model is a "thinking" variant; reasoning tokens were eating the output budget. Setting the thinking budget to zero restored full answers.
• Cost under load. The free tier dried up during a 50-question QA run, so generation falls back through a cost-ordered chain of models to stay within quota.

Isolating the demo (a security note)

The public demo and the real product shared one API key, so demo traffic burned the production quota. I added product-aware key selection in the edge function — the demo tenant uses its own key — so a curious visitor can never exhaust or bill the real one.

Screenshots

Screenshots are from the public white-label demo (fictional "VoltLine" energy tenant), not a client deployment.

Wingman — answering a live price objection	Pitch kit — tools, demos & key numbers

Stack

Supabase (pgvector · Edge Functions) · Gemini · RAG · TypeScript / React · Cloudflare Pages

Live demo & the technique

• Live demo: https://lucky-cat.pages.dev/pitcher-template/
• The retrieval pipeline, runnable end-to-end: github.com/augbastos/rag-demo

More work: augustobastos.pages.dev