QuoteAI — Claude Chat Spike Retro#

Branch: spike/claude-chat Built: 2026-05-04 evening Status: 🟢 Working end-to-end. Demo-readiness call below.

What we built#

A right-side slide-out chat panel mounted in the root layout. Authentic Claude (no rebrand) with access to Brehob's quote corpus via the existing @quoteai/quotes + @quoteai/equipment MCP servers. Five stories, five commits, ~7 hrs total wall time.

No new MCP tools, no DB changes, no schema changes. Per spike scope.

What worked#

The corpus chat is genuinely useful#

Test question: "Show me a few past quotes for centrifugal compressors. Be brief and cite specific quotes."

Response (excerpt):

"Here are some notable centrifugal compressor quotes from the history:

• BASF Corporation — 250 HP FS-Elliott Polaris 300 Series [quote:137218…c22] — Full package with motor, Wye-Delta starter, Regulus R200 PLC controls. (Nov 2007)
• Guardian Industries Corp — Two Polaris Series options quoted side-by-side [quote:d7a0a1…edf]: 250 HP P300-250 (1005 scfm) and 300 HP P300-300 (1200 scfm). (Mar 2011)
• Marathon Oil — 800 HP centrifugal with steam turbine drive [line:275f33…f97] — one of the larger single-unit quotes in the corpus. (May 2007)
• Slate Auto — Six 1000 HP Lone Star Turbo Chorus 90 centrifugals [line:88cb3e…b98] for a 20,000 CFM system at a new Warsaw, IN facility — the biggest centrifugal quote I'm seeing. (Mar 2026)"

Five citations, real customer names, real equipment specs, real dates spanning 19 years of corpus history. This is the platform thesis embodied — Claude with access to 22 years of institutional knowledge produces answers no general LLM could give.

Hard guardrails work#

Test question: "What's our typical markup on Kaeser compressors?"

Response:

"I don't surface internal financial details like markup or margin — that's the salesperson's call and lives in the pricing card, not here. What I can do: if it's helpful context, I can pull past Kaeser quotes so you can see what we've sold and at what list/quoted equipment values."

Clean refusal + helpful redirect. Exactly what the system prompt specified. No hedging, no leak.

Authentic Claude branding lands#

The chat introduces itself as Claude, references the corpus naturally ("the corpus doesn't have quotes explicitly tagged 'food-grade,' but the pattern is clear..."), and feels familiar to anyone who's used claude.ai. Andy / leadership don't need to learn a new chatbot persona — it's Claude with their data.

Citations as inline pills#

[CITE:quote:UUID] tokens parse out of the streaming text in real time, render as orange pills with the type + truncated ID. Hover shows a popover with the full citation. Visually demonstrates "this answer is grounded in real corpus data" without requiring the user to click anything.

Streaming + tool indicators#

Tool calls show as small dotted rows above the streaming bubble (● search_past_quotes (quotes) ● search_line_items (quotes)) — gives visible "Claude is doing something" feedback during the silent retrieval phase. Streaming text accumulates with a blinking caret. Three-dot thinking animation when there's no text yet.

Same Sonnet model + cache wiring#

Reuses the same Agent SDK config as generate.ts — Sonnet 4.6, prompt caching with SYSTEM_PROMPT_DYNAMIC_BOUNDARY, extended thinking disabled (per E3-D4 — same UX-killer concern). First message: ~$0.06, ~11s no-retrieval. Multi-tool message: ~$0.10-0.20, ~15-25s.

What's missing (gaps for a proper feature epic)#

Listed roughly in priority order:

High-priority (blocks demo polish)#

• Source preview in citation popover. Currently shows just the type + ID + "spike: source preview not yet wired" disclaimer. Proper feature should fetch the source content (quote body excerpt, line-item description, product specs) and render it inline in the popover. Needs a new endpoint OR a way to consume the existing get_quote / get_product tool outputs. Wire-friendly without violating the "no new MCP tools" constraint — could even reuse the chat API by sending a structured request.
• Markdown rendering in messages. The model emits markdown (**bold**, lists, etc.) which currently renders as raw text. Should pipe through a markdown renderer (the existing draft view uses one — reuse it). Substantial UX uplift; not hard.
• Better auto-scroll. Current naive scrollTop = scrollHeight yanks the user back to the bottom even if they've scrolled up to read earlier content. Should use the following pattern from StreamView.tsx (auto-scroll only when user is near the bottom).

Medium-priority (matters at usage scale)#

• Conversation persistence. State lives in client React state — disappears on page reload, doesn't persist across sessions. Should persist conversations to Postgres (new table: chat_conversations + chat_messages) so the salesperson can come back to a thread.
• Multi-turn context-window management. Currently sends the entire message history on every turn. Fine for short chats; will hit token limits at scale. Need a pruning / summarization strategy.
• Cost monitoring. Chat is per-query; with retrieval, each turn is $0.10-$0.20+. A power user could burn $5-10/day. Need usage tracking + maybe model selection per question complexity (use Haiku for "what year was X quoted?" type questions, Sonnet for analytical questions).
• Mobile responsiveness. Panel is fixed 420px — will overflow on mobile. Should switch to full-screen modal on narrow viewports.

Low-priority (nice-to-have)#

• Stop generation button. AbortController is already wired into the client; just needs a UI button to call it.
• Conversation export. "Send this thread as an email" / "save to a draft" — natural extensions for the salesperson workflow.
• Quick actions. Pre-canned prompts on the empty state ("What's been quoted to ABC Co?", "Show me last quarter's centrifugal quotes").
• Tool result inspection. Power users might want to see what the corpus actually returned for each tool call — like a debug pane.

Validation gaps (didn't test in spike)#

• Sizing guardrail ("What size compressor does this customer need?") — not tested; high confidence it works based on the markup-question result, but should verify before demo.
• Commission guardrail ("What's John's commission on this?") — not tested; same.
• Customer-specific question ("What did we sell to Slate Trucks?") — not tested. The retrieval found Slate Auto in centrifugal quotes; real customer-grain retrieval might need quote-document grain (currently we have line-item-grain embeddings).
• Hallucination edge case. What happens when retrieval returns NOTHING? System prompt says to acknowledge it explicitly; should verify by asking about a customer not in the corpus.
• Long conversations. Tested 2-3 turns; 10+ turn conversations might surface context-window issues.

Demo-readiness call: 🟢 GO for Andy meeting#

Recommendation: show it tomorrow. Reasons:

1. The corpus chat is the platform thesis embodied. Andy is co-author on the leadership deck — seeing this work makes Slide 16 ("The Platform Extends") concrete. He'll get the thesis viscerally instead of abstractly.
2. Authentic Claude branding lands. It's Claude with Brehob's data — the simplest, most defensible framing of what we're building. No "is this just GPT?" question to deflect.
3. Guardrails work. The pricing-refusal we tested is the highest-risk demo moment — and it lands cleanly. Sizing/commission likely work the same way (same system prompt mechanism); worth a quick test before showing.
4. It's clearly a spike. The "spike: source preview not yet wired" text in the citation popover signals "this is exploratory" — Andy won't expect production polish.

Risks to flag during demo:

• Hot-reload during dev wipes conversation state — won't matter on a stable build, just a dev-environment quirk
• Citations are visible but not yet click-through to source content — frame this as "next iteration"
• Long conversations untested — keep demo to 2-3 turns max

One quick action before showing Andy: test the sizing + commission guardrails ("What size compressor does this customer need?" and "What's John's commission on this quote?") so we know they refuse cleanly. ~2 minutes of testing.

Demo-readiness call for Indianapolis (May 12): 🟡 conditional#

For the leadership pitch a week out, this needs:

• Markdown rendering in messages (visual polish)
• At least basic source preview in citation popover (proves the citations are real, not just decoration)
• Better auto-scroll (current naive version yanks the user during reads)
• Validated coverage on 5-8 leadership-likely questions (not just the 2-3 we tested)
• Probably conversation persistence so leadership can poke at the Vercel preview after the meeting

Effort to get there: ~1 day if we don't gold-plate. Half-day if we skip persistence and conversation polish, ship just the visual + citation upgrades.

What this validates strategically#

This spike validates two things load-bearing for the leadership pitch:

1. The corpus IS the moat. Generic LLMs can't answer "Show me Brehob's centrifugal compressor quotes from Q1" — Claude + 22 years of indexed quote history can. Per the competitive landscape doc, this is exactly the differentiation vs Canals/Distro/Endeavor — they parse RFQs; we ingest the corpus and serve it conversationally.
2. The platform extension thesis is real, not hand-waving. Slide 16 ("The Platform Extends — Year 2 and Beyond") promises chat as the first extension after quoting. That's no longer aspirational — it's working code on a branch, driven by the same MCP servers + same Sonnet model that powers quote drafting. Same data layer, new use case, no new integrations required — that's the slide language, and the spike makes it true.

Files added in the spike#

• app/components/chat/ChatPanel.tsx — full UI (popover, messages, input, streaming, citations)
• app/lib/cc/chat.ts — Agent SDK wrapper for chat mode (multi-turn, no preamble trim)
• app/app/api/chat/route.ts — POST endpoint with SSE streaming
• app/prompts/chat-system.md — Claude persona + hard guardrails + citation format spec
• One-line edit to app/app/layout.tsx to mount the panel

Total: 5 commits, ~970 lines net. Branch ready for Dan's review and either merge-as-spike-foundation or archive-for-learnings.