Platform update
Vehicle Context Engine — Unified Intelligence for Every Widget
One vehicle context packet now powers every Carvia widget — reconciling identity, inventory, valuation, history, and specs in a single grounded retrieval pass so AI insights stay consistent, accurate, and trustworthy on your VDP.
Release Overview
Every Carvia widget on your VDP — insights, valuation, score, chat — now draws from a single, unified intelligence packet built for that exact vehicle before any AI generates a word.
The Vehicle Context Engine is the platform layer behind smarter widget content. Instead of each widget fetching its own data and assembling its own prompts independently, Carvia now retrieves identity, inventory, valuation, history, specs, and market position in one coordinated pass — normalizes conflicts, scores data quality, and feeds a structured context packet to AI generation.
For dealers, the result is subtle but critical: AI insights that agree with your listing price, your inventory feed, and your valuation widget. No more shoppers seeing a “2020 Camry SE” in one tab and a different trim in another. No more AI inventing title problems that are not in the vehicle history. One source of truth, one retrieval pass, one generation pipeline.
The Problem We Solved
Before the Vehicle Context Engine, Carvia widgets were powerful — but they were also fragmented.
Fragmented data retrieval
Each widget tab — Insights, Valuation, Score, History — pulled data on its own schedule. VIN decode might resolve one trim while your inventory feed showed another. Valuation used listing price from the embed; insights used price from a different source. The widgets looked like one product, but under the hood they were four independent systems guessing at the same vehicle.
Inconsistent AI outputs
AI insights were assembled from piecemeal prompt fragments — whatever data happened to be available when that specific widget loaded. Without a single authoritative context, the model could confidently state facts that contradicted your valuation widget or your NMVTIS history report. Shoppers notice. Trust erodes fast when the price says “fair deal” but the insights tab warns about problems that do not exist.
Inventory feed drift
Dealers live with imperfect data. A trim mismatch between your DMS export and VIN decode is routine. A special price that updated overnight but has not propagated to every system is common. Without explicit conflict detection, these gaps surfaced as confusing widget content — or worse, as AI hallucinations filling in the blanks.
No cache discipline
When listing price drops $2,000 on a Saturday, insights generated Friday afternoon should not still be cached as if nothing changed. The old architecture lacked a deterministic way to know when vehicle context had materially changed — leading to stale AI content on live inventory.
What's Inside the Context Packet
The VehicleContextPacket (v1) is a versioned, cacheable JSON structure. Every field is normalized before it reaches AI. Here is what each section contains — and why your dealership should care.
Canonical vehicle identity
Year, make, model, trim, body style, drivetrain, engine, fuel type, and transmission — reconciled from VIN decode and your inventory feed. When the two sources disagree, conflicts are recorded with severity levels. The canonical identity is what every widget displays and what AI references. No more “2021 Honda Accord” in insights and “2021 Honda Accord Sport” in specs.
Your live listing data
Stock number, status, new/used flag, mileage (from embed request or inventory row), list price, special price, effective listing price, photo count, VDP URL, and import freshness timestamp. Listing price resolution follows a clear priority: embed request → special price → list price. When your weekend sale price hits the feed, the packet reflects it on the next generation.
Fair value summary
Dealer fair value, private party value, price range, confidence score, strategy, and cache status — summarized from your Valuation Orchestration pipeline. The packet knows whether valuation data is available and whether it came from cache. AI insights reference fair value only when both listing price and dealer fair value are present — preventing invented “great deal” or “overpriced” claims.
Title, accident, and recall signals
Title brand flags, odometer issues, accident/salvage signals, ownership history indicators, and open recall count — summarized from NMVTIS data. AI is explicitly instructed: do not claim title brands, accidents, salvage, open recalls, or odometer issues unless they appear in this section. This is the guardrail that stops hallucinated red flags.
Build data and equipment
MSRP, options summary, EV-specific details, and safety summary from EVD2 build data. Powers accurate feature callouts in insights and supports EV Insights content for electric inventory.
Price position vs. fair value
Whether your listing price sits below, at, or above the Carvia fair value estimate — computed with a 3% tolerance band. This gives AI a grounded anchor for value framing without inventing market comparisons. Comparable vehicle retrieval for market context is on the roadmap (see below).
Completeness score and warnings
A 0–100 completeness score based on seven core fields: identity year/make/model, inventory found, valuation available, history source, and specs source. Missing fields and identity conflicts are listed explicitly. When completeness is low, AI acknowledges uncertainty rather than filling gaps with guesses.
Deterministic invalidation
A context input hash computed from VIN, client ID, mileage, ZIP, listing price, inventory timestamp, and valuation fingerprint. Insights regenerate only when something material changes — not on every page load. Invalidation signals track inventory and valuation update timestamps for precise cache busting.
What's New
VehicleContextPacket (v1)
Standardized, versioned intelligence structure shared across the widget platform. Built by VehicleContextBuilder on every context-aware widget load. Exposed via internal API at /internal/vehicle-context/[vin] for downstream services.
Context-based AI insights generation
The Insights widget can now generate from the full context packet via /internal/vehicle-insights/from-context — replacing piecemeal prompt assembly with structured vehicle intelligence. The AI agent receives the entire packet as JSON and is instructed that it is the only authoritative source of vehicle-specific facts.
Grounded generation guardrails
System instructions enforce buyer-friendly, non-alarmist framing aligned with Buyer-Fit Insights: “Who is this a good fit for?” not “Should you buy?” Banned terms block fear-based language. Output budgets cap summary and paragraph lengths for scannable VDP content. New-car detection (<1,000 miles) shifts focus to fit and features instead of used-car inspection alarmism.
Identity conflict detection
When decoded VIN data disagrees with your inventory feed on year, make, model, trim, or body style, conflicts are flagged with severity levels (info or warning). The canonical identity resolves intelligently — but the conflict is preserved in the packet so AI can acknowledge uncertainty instead of picking a side silently.
Shadow mode rollout
Context-based insights deploy in three modes: disabled (legacy path only), shadow (both paths run; context output compared but legacy still serves shoppers), and primary (context path serves shoppers). Shadow mode generates comparison records for quality audit before cutover — so your shoppers never see experimental output until it is proven equal or better.
Context-input hashing and smart cache
Deterministic cache keys from packet inputs mean insights regenerate only when inventory price changes, valuation updates, or request parameters shift. No wasted AI calls on unchanged vehicles. No stale insights on vehicles whose price dropped this morning.
Prefetch optimization
The builder accepts prefetched vehicle data and valuation from the widget carousel load — avoiding redundant external API calls. Data retrieved once for the carousel is reused to build the context packet, keeping VDP load times fast even as intelligence depth increases.
Share lead identity resolution
When shoppers tap Get My Full Report, vehicle identity for the share lead resolves from the same carousel data the context engine normalizes — preferring ownership cost and specs over insights placeholders that may lag. Accurate year/make/model on every shared report, even when the insights tab is still loading.
How It Works
Six steps from VDP load to grounded AI insights — all behind the scenes.
- Request arrives. The widget embed sends VIN, mileage, ZIP code, listing price, and client ID from your VDP. Inputs are normalized (mileage bounds, price formatting, ZIP validation) before any data retrieval begins.
- Inventory resolve. Carvia loads the active inventory row for your client/VIN pair — stock number, prices, photos, status, and import timestamp. Listing price is resolved with clear source priority: embed request → special price → list price.
- Data prefetch. VIN decode (EVD2), NMVTIS history, and valuation data are loaded from cache when available. The builder reuses data already fetched for the widget carousel — no duplicate external calls on a typical VDP load.
- Packet build. Identity is reconciled from decode and inventory snapshots. Conflicts are detected. History, specs, valuation, and market position are summarized into structured sections. Data quality is scored. A deterministic context input hash is computed for caching.
- Insights generation. The full packet is sent to the AI agent with strict grounding rules: the JSON is the only authoritative source. The model generates buyer-fit insight sections — fit, ownership outlook, inspection priorities, market context — within output budgets and banned-term guardrails.
- Widget render. Insights display in the Widget Carousel. The same context packet architecture is designed to power Vehicle Report Chat and future widget paths with identical data consistency.
allowExternalFetch: false). External NMVTIS, EVD2, and valuation API calls happen during the initial carousel data fetch — not again during packet build. One retrieval pass, shared across widgets.
What Shoppers Experience
Shoppers never see “Vehicle Context Engine.” They see better content.
Consistent vehicle identity
Every tab in the widget carousel references the same year, make, model, and trim. When a shopper moves from Insights to Valuation to Score, they are looking at one vehicle — not four interpretations of a VIN.
Grounded, trustworthy insights
AI insights reference only facts present in your vehicle data. Title brands appear when NMVTIS reports them — not because the model guessed. Fair value comparisons appear when both listing price and valuation are available — not because the model assumed a deal. When data is incomplete, insights acknowledge uncertainty instead of inventing details.
Buyer-fit framing, not fear
Insights answer “Who is this a good fit for?” and “What should you confirm before signing?” — aligned with the Buyer-Fit shift from February 2026. No alarmist language. No “RUN AWAY FROM THIS CAR” headlines. Consultative guidance that builds confidence on your VDP.
Accurate shared reports
When a shopper taps Get My Full Report, the shared vehicle page carries the same normalized identity the context engine resolved — correct photos, correct trim, correct price. Your salesperson sends a report that matches what the shopper saw on the VDP.
Faster repeat visits
Smart caching means returning shoppers see fresh insights when your price changes — but do not wait for redundant AI regeneration when nothing has changed. The VDP feels responsive even on data-rich widget loads.
Value by Role
For general managers
Trust in your VDP technology translates to trust in your dealership.
- Data consistency — widgets agree with each other and with your inventory feed
- Reduced liability — AI cannot invent title problems or recall warnings not in your vehicle history
- Platform foundation — every future widget improvement builds on one intelligence layer
- Quality-controlled rollout — shadow mode proves improvements before shoppers see them
Scenario: A shopper mentions a “salvage title warning” on a clean-title vehicle. Before the context engine, this could happen when AI hallucinated from incomplete prompts. Now, title brands come only from NMVTIS data in the packet — and your BDC can confidently correct the record.
For marketing and internet directors
Your VDP is your highest-intent digital real estate. Content quality matters.
- Brand-safe AI — banned terms and output budgets prevent off-brand or alarmist copy on your VDP
- Price-change responsiveness — sale prices propagate to AI insights on the next cache invalidation
- New and used parity — context engine detects new vehicles and shifts insight framing appropriately
- Measurable improvement — shadow comparison records quantify insight quality gains over legacy generation
Scenario: You run a weekend sale dropping prices on 40 units. Monday morning, refreshed insights reference the new special prices — not stale list prices from before the event. Shoppers see current value framing.
For sales professionals
Sell with confidence knowing the widget matches the vehicle on the lot.
- Accurate shared reports — Get My Full Report carries the same identity the VDP displayed
- Coaching-ready insights — buyer-fit framing gives conversation starters, not scare tactics
- Valuation alignment — insights reference the same fair value your Valuation tab shows
- Trim accuracy — identity reconciliation catches feed mismatches before shoppers do
Scenario: You walk a customer through the Insights tab on the lot. The fit summary, ownership outlook, and value positioning all reference the same vehicle data you see in your inventory system. No awkward moments when the widget says something different from your window sticker.
For inventory and data teams
Identity conflict detection surfaces trim and YMM mismatches between your DMS export and VIN decode — giving your data team actionable flags to clean feed quality. Data quality scores and missing-field warnings help prioritize which inventory rows need attention before they reach shoppers.
On the Roadmap
The Vehicle Context Engine is stage 1 of a broader grounded intelligence platform. What ships today is production-ready and delivering value. What comes next extends the packet with deeper retrieval.
External knowledge-base RAG (not yet live)
The context packet includes a knowledgeRetrieval field — currently disabled with reason not_implemented_stage_1. This is the hook for curated external knowledge: repair guides, technical service bulletins, regional recall advisories, and model-specific owner community insights. When live, RAG will augment structured vehicle data with retrieved documents — not replace it.
Comparable vehicle retrieval in market context (not yet live)
The market section currently computes price position vs. fair value from listing price and valuation data. Comparable vehicle cohort retrieval — the same retrieval layer powering Valuation Orchestration — will feed directly into the context packet for richer market positioning in AI insights.
Primary cutover for all clients
Context-based insights are rolling out per client via feature flags and environment mode (disabled → shadow → primary). Shadow mode comparison records validate quality before primary cutover. Contact your Carvia account team to confirm your rollout status.
Release Timeline
/internal/vehicle-insights/from-context endpoint, grounded prompt engineering, and shadow mode orchestration.Getting Started
The Vehicle Context Engine runs behind the scenes — no dealer configuration required for the core packet build.
- Confirm your widget embed is current. The context engine activates automatically on supported widget carousel builds. Your Carvia account team can verify your embed version includes context-aware insights.
- Ask about your insights rollout mode. Context-based insights may be in shadow (quality comparison, legacy still serves) or primary (context path live) depending on your account. Either way, the packet build and data normalization are active.
- Keep your inventory feed clean. Identity conflict detection works best when your DMS export matches VIN decode. Trim mismatches are flagged — fixing them improves insight accuracy and share report quality.
- Update prices promptly. The context engine invalidates cached insights when inventory timestamps or listing prices change. Fresh prices in your feed mean fresh value framing on your VDP.
- Pair with analytics. Use Widget Analytics & Impact Reports to measure whether improved insight quality translates to higher engagement and share leads.
- Send shared reports with confidence. Get My Full Report resolves vehicle identity from the same normalized context — accurate YMM on every outbound report.
The Vehicle Context Engine is the intelligence layer that makes every Carvia widget on your VDP speak with one voice — grounded in your data, consistent for shoppers, and built to get smarter with every release.