Konsil Deal Diagnostic

Konsil is a pattern-diagnosis tool for complex B2B sales deals, grounded in a proprietary sales methodology ("How Strategic Sellers Build Deals That Don't Exist Yet"). Sellers describe a stuck deal; the system identifies which pattern applies — Mode 2 Ceiling, Underpowered Champion, Thesis Without a Sponsor, Currency Mismatch, Federation Map problem — and returns three strategic moves, each with a specific tradeoff and an exact chapter citation from the methodology.

AI Architecture — Hybrid RAG Pipeline

1. Query embedding — The deal situation is embedded via Google Gemini Embedding API (768-dim vectors, RETRIEVAL_QUERY task type).
2. Two-tier retrieval (Supabase PostgreSQL + pgvector):
   • Foundational chunks — Core framework chapters (Modes, Currencies, Roles, Relationship Types, Pulling Levers) always retrieved first to anchor every response in the methodology.
   • Situational chunks — Hybrid search: vector cosine similarity + BM25 full-text search, combined via Reciprocal Rank Fusion (k=60, top 12 results). Ch30-34 (AI-only chapters) explicitly excluded to prevent bleed into deal-strategy advice.
3. Generation — Gemini 3 Flash Preview — Temperature 0.3 for consistent output. System prompt enforces citation-first JSON output: each move must name a specific tradeoff cost (time, personal currency, timeline extension, deal size) — not abstract difficulty. Banned terminology list enforced in both the prompt and a post-processing sanitizer (regex replacements catch phrases the model generates despite instructions).

Prompt Engineering & Quality Control

Every response is validated by a 20-scenario test suite (test_oracle.py) covering all chapters. Tests check for citation accuracy, banned terms, weak tradeoffs, AI chapter bleed, and chapter numbering correctness. Deviation audits flag when the model paraphrases instead of citing directly.

Voice Input

Web Speech API (SpeechRecognition / webkitSpeechRecognition cross-browser fallback). Handles the Android resultIndex=0 bug (duplicated final results) by tracking lastFinalIndex. Only the last interim result is appended to prevent "the the the" artifacts from rapid interim events. Animated mic button with pulsing ring indicator.

Data Ingestion

A one-time Python script (scripts/ingest.py) parses the methodology DOCX: detects heading hierarchy (H1 for Ch1-5, H2 for Ch6+), splits into 80-600 word chunks, embeds each via Gemini, and loads into Supabase with IVFFlat vector indexing (50 lists for 768-dim embeddings).

Deployment

Vercel (Next.js 14 + React 18 + TypeScript). Supabase PostgreSQL with pgvector for hybrid semantic + full-text search. Web only — responsive dark-theme UI works on desktop and mobile browsers. No native app.