Platform update
Valuation Orchestration — Comparable Retrieval & AI Routing
Grounded fair-value estimates powered by comparable retrieval, intelligent ML/LLM routing, comp-blend anchoring, and multi-factor confidence — so every price on your VDP reflects real market signals.
Release Overview
Every fair-value estimate on your VDP is now built the way modern grounded AI systems work: retrieve market signals first, then reason — not the other way around.
Valuation Orchestration replaces single-model AI guessing with a structured pipeline. Comparable vehicles are retrieved and scored for similarity. Confidence is computed from data completeness, market quality, and risk signals. An orchestrator routes between Vertex ML models, structured LLM reasoning, and comp-anchor blending. The final dealer fair value and private party range are anchored to retrieved market data — with full audit telemetry on every run.
For dealers, this means valuations shoppers actually trust. A price that reflects local market context, vehicle history, mileage, and comparable signals — not a black-box number that drifts from what buyers see on Cars.com next door.
The Problem We Solved
A confident-sounding wrong price is worse than no price at all.
The single-model problem
Before orchestration, Carvia valuations relied on AI models reasoning in isolation — fed vehicle data and asked to produce a fair value. Without comparable anchoring, models could produce prices that sounded authoritative but did not match local market reality. A shopper comparing your VDP valuation to three similar listings on AutoTrader would spot the disconnect immediately. Trust evaporates.
The geography problem
National average pricing ignores the market your shopper lives in. A truck worth $38,000 in rural Texas is not worth $38,000 in downtown Toronto. Market context (March 2026) solved localization — but localization without retrieval still left AI reasoning without an anchor to local comparable inventory.
The history-blind problem
A salvage-title vehicle and a clean-title vehicle of the same year/make/model should not receive the same fair value. Without explicit risk dampening in the valuation pipeline, history signals could be underweighted — producing optimistic prices on vehicles that savvy shoppers would price far lower.
The opacity problem
When a valuation goes wrong, there was no audit trail. No record of which model ran, whether comparables were found, what blend strategy was used, or why confidence was low. Debugging quality issues and improving over time required guesswork.
What's New
Comparable retrieval service
Builds a scored cohort of comparable vehicles matched on year/make/model, ZIP, mileage, and market context. Each comparable is ranked by a composite score: 55% similarity (trim, mileage proximity, ZIP region), 15% freshness (listing recency with 45-day half-life), and 30% distance (geographic proximity). The service computes median price, price range, high-quality comparable count, and a retrieval quality score — all before any AI call runs.
AI orchestration and routing
A pure routing layer — ValuationAiOrchestrator — decides the execution plan for every valuation: Vertex ML primary with LLM structured follow-up, LLM-only with comp blend, hybrid model plus comp anchor, or LLM primary with numeric shadow evaluation. Routing depends on Vertex eligibility, runtime mode, data completeness, and whether a numeric model provider is available. Every decision is logged with reason codes for audit.
Comp-blend anchoring
AI-generated fair values are blended with comparable median prices using confidence-weighted strategies. Three blend modes: comparables primary (strong market data dominates), hybrid (weighted mix of AI and comps), and AI primary (weak comp data, AI leads). Salvage history, flood records, odometer issues, and excessive title brands automatically reduce comp confidence weight — preventing over-anchoring on unreliable market signals for risky vehicles.
Google Search grounding on valuation LLM
Structured valuation LLM calls can use Google Search retrieval for real-time market grounding — pulling current listing trends and regional pricing signals into the reasoning context. Enabled by default on the Vertex Gemini path; configurable per environment. When Vertex is unavailable, the system falls back to direct Gemini with the same structured JSON output format.
Multi-factor confidence scoring
Every valuation receives a 0–100 confidence score and label (high, medium, low) computed from: data completeness (NMVTIS, EVD2, VIN decode), comparable retrieval quality, high-quality comp ratio, price spread tightness, AI-vs-comp agreement, and risk flag penalties. Low-confidence valuations are flagged internally — never shown as alarming warnings to shoppers, but available for quality monitoring.
Valuation run persistence
Full run snapshots stored in vehicle_valuation_runs — execution path, blend strategy, comparable count, confidence breakdown, Vertex ML outcome, LLM provider, Google Search enabled flag, and orchestration plan kind. Powers offline evaluation, ML training pipelines, shadow comparison, and promotion workflows. Every valuation is auditable.
Vertex ML shadow evaluation
In shadow mode, the LLM path serves shoppers while Vertex ML runs in parallel for comparison. Shadow results are logged with tier classification and summary — enabling safe model promotion without affecting live valuations. Production cutover happens only when shadow quality meets thresholds.
New-car valuation path
Dedicated prompts and system instructions for new vehicles — MSRP context, depreciation framing, and market positioning vs. used alternatives. Orchestration detects new-car signals (low mileage, no registration history) and routes to the appropriate prompt path automatically.
How It Works
Seven steps from VIN to grounded fair value — every one logged.
- Feature build. VIN decode (EVD2), NMVTIS history, and market context from the shopper's ZIP or postal code are normalized into valuation features — year, make, model, trim, mileage, history flags, data completeness score, and regional market parameters.
- Comparable retrieval. A scored cohort of comparable vehicles is assembled for the subject's YMM, ZIP, and mileage. Comparables are filtered, ranked by composite similarity score, and summarized into median price, price range, retrieval quality score, and high-quality count.
- Pre-AI strategy selection. Based on comparable availability and data completeness, the engine selects a strategy: ai_with_market_context (comps available), ai_with_features (full vehicle data, no comps), or ai_only (minimal data fallback).
- Orchestration routing. The orchestrator evaluates Vertex ML eligibility and runtime mode, then builds an execution plan — which steps to run (numeric primary, LLM structured, comp anchor blend, numeric shadow eval) and in what order.
- AI generation. The selected path runs: Vertex ML prediction, structured LLM with optional Google Search grounding, or both in sequence. Output is structured JSON — dealer fair value, private party value, price ranges, and market analysis narrative.
- Comp blend and confidence. AI midpoint is blended with comparable median using the effective comp confidence weight (risk-dampened). Blend strategy is selected (comparables primary, hybrid, or AI primary). Final confidence score and label are computed from the full signal set.
- Widget display and persistence. The Carvia Valuation widget renders dealer fair value, private party range, and market analysis on your VDP. The full run snapshot is persisted for audit. The same valuation summary feeds into the Vehicle Context Engine packet and Carvia Score calculation.
model_only,
llm_only,
hybrid_model_then_llm,
hybrid_model_plus_comp_anchor,
llm_primary_numeric_shadow,
fallback_only.
Shoppers see the final price — not these internals.
Confidence & Comp Blending
The two mechanisms that keep valuations honest — explained without the math homework.
How comp blending works
After AI produces a fair value estimate, the engine computes comparable statistics: weighted average price, median price, and price range from the top-ranked cohort. A comp confidence weight (0–1) is calculated from retrieval quality, high-quality comparable ratio, mean similarity, and price tightness. History risk flags (salvage, flood, odometer issues) dampen this weight.
The blend strategy then decides how much to trust comps vs. AI:
- Comparables primary — strong comp data (5+ high-quality comps, quality score ≥ 0.45, weight ≥ 0.62). Market median anchors the final price.
- Hybrid — moderate comp confidence. AI and comps contribute proportionally.
- AI primary — weak or missing comp data. AI estimate leads with minimal comp influence.
- Fallback — retrieval failed or heuristic path used. Lowest confidence tier.
How confidence scoring works
The final confidence score (0–100) combines four signal groups:
- Data completeness — NMVTIS, EVD2, and VIN decode presence (up to 32% of score)
- Comparable quality — retrieval quality, average final score, high-quality ratio (up to 38%)
- Price consistency — tight comparable price spread signals reliable market data (up to 12%)
- AI vs. comp agreement — when AI and comp median align, confidence rises; divergence lowers it (up to 18%)
Risk penalties subtract for salvage/total loss (−12%), rebuilt/flood (−6%), multiple title brands, odometer issues (−8%), and extreme mileage (−5%). Labels: high (≥75%), medium (45–74%), low (<45%).
What Shoppers See
A fair, localized price they can compare to your window sticker — not a generic national estimate.
Dealer fair value and range
The valuation widget displays a dealer fair value with a low–high range, plus a private party estimate. Numbers are scaled consistently — dealer and private party bands move together when comp blending adjusts the midpoint. Shoppers see one coherent story, not conflicting figures.
Listing price positioning
When your listing price and the Carvia fair value are both available, the widget and Vehicle Context Engine compute price position — below, at, or above fair value (3% tolerance band). AI insights reference this positioning only when both numbers are present — no invented “great deal” claims on vehicles priced above market.
Market analysis narrative
A concise market analysis explains the valuation in shopper-friendly language — grounded in retrieved comparable signals and vehicle-specific data. New vehicles get fit-and-value framing; used vehicles get history-aware context. Banned terms and output budgets keep copy professional and non-alarmist.
Downstream impact on Carvia Score
Valuation quality feeds directly into the Carvia Score on your VDP and SRP Listing Badge. A well-priced, clean-history vehicle with strong market anchoring scores higher. Orchestration ensures the score reflects grounded data — not AI optimism.
Value by Role
For general managers
Pricing credibility protects your brand and your gross.
- Defensible valuations — every price is retrieval-anchored with audit trail
- Risk-aware pricing — branded-title and high-risk vehicles valued conservatively
- Continuous improvement — run persistence enables model quality monitoring and promotion
- Score integrity — Carvia Score reflects orchestrated valuation, not guesswork
Scenario: A shopper challenges your price on a used F-150. Your salesperson opens the valuation widget showing a fair value range anchored to local comparable trucks — not a national average. The conversation shifts from “you're overpriced” to “here's how this truck compares to the market.”
For marketing and internet directors
Transparent pricing is your competitive moat on the VDP.
- Local market accuracy — valuations reflect ZIP-level context, not generic national data
- Canadian readiness — market context and orchestration support cross-border dealer groups
- Consistent VDP story — valuation, score, and insights reference the same fair value
- Google Search grounding — real-time market signals supplement structured comparable data
Scenario: Your competitor runs a weekend sale on trucks. Google Search grounding helps Carvia valuations reflect current market movement — so your fair value ranges stay current even between inventory import cycles.
For sales professionals
Arm your team with a number they can stand behind.
- Negotiation anchor — fair value range gives a data-backed starting point for price discussions
- Below-market listings stand out — when your price is below fair value, the widget shows it — a conversion tool, not just information
- History-aware framing — valuations on clean-title vehicles are stronger; your team knows which units to lead with
- Shared report consistency — Get My Full Report carries the same valuation the VDP displayed
Scenario: A customer says they found the same vehicle cheaper online. Your salesperson pulls up the Carvia valuation showing private party range and comparable market position — shifting the conversation from price alone to total value, history, and fit.
For F&I and desk managers
Accurate fair value reduces surprises at the desk. When trade-in discussions reference Carvia valuations shoppers already saw on the VDP, the numbers align — fewer “that's not what your website said” moments and smoother deal structuring.
On the Roadmap
Valuation Orchestration is production-ready and delivering grounded valuations today. Two layers are scaffolded for future releases:
Live listing feed for comparable retrieval (in progress)
Comparable retrieval currently builds scored cohorts using an internal synthetic benchmark strategy — full scoring pipeline (similarity, freshness, distance) is live, with the architecture ready to swap in a real listing feed when available. When live listings connect, comp-blend anchoring becomes even more precise because median prices reflect actual market inventory, not benchmarks.
External knowledge-base RAG (scaffolded, not live)
The orchestration pipeline includes hooks for rag_knowledge_retrieval — retrieving curated automotive reference documents (repair guides, TSB databases, regional advisories) before LLM reasoning. The current provider is a no-op (NoOpValuationKnowledgeRetrievalProvider) with skip reason phase3_placeholder_not_implemented. Today's grounded valuations use comparable retrieval and Google Search grounding — not vector RAG over proprietary knowledge bases.
Release Timeline
Getting Started
Valuation Orchestration runs automatically on every Carvia Valuation widget request. No dealer configuration required.
- Confirm your valuation widget is active. Orchestration powers the Valuation tab in your Widget Carousel embed. If the carousel is live on your VDPs, orchestrated valuations are already running.
- Verify market context. Ensure your dealer domain is registered in the Carvia client portal for automatic ZIP/postal resolution. Valuations without location context lose comparable precision.
- Keep inventory feeds current. Listing price flows into valuation features and price-position calculations. Stale prices in your feed mean stale fair-value comparisons on the VDP.
- Review valuation on your VDP. Open any vehicle on your site and check the Valuation tab. Dealer fair value, private party range, and market analysis should reflect the vehicle's year, mileage, history, and local market.
- Pair with SRP badge. If you have the SRP Listing Badge, orchestrated valuations also power the Carvia Score on your search results page.
- Track engagement. Use Widget Analytics to see how often shoppers engage with the Valuation tab — and whether valuation engagement correlates with share leads and outcomes.
Valuation Orchestration turns every fair-value estimate on your VDP into a retrieval-anchored, confidence-scored, auditable price — the kind of transparency that earns shopper trust and supports your team at the desk.