NHTSA TSB & Manufacturer Communication Monitor

Monitor official NHTSA TSB data and manufacturer communications for new or revised technical service bulletins, vehicle service campaigns, and warranty/service updates. Turn NHTSA's bulk archive into review-ready signals for fleet risk monitoring without repeatedly downloading, parsing, filtering, and comparing the source yourself.

Each paid result represents one matching document + make + model + model-year identity. It includes the verbatim NHTSA summary, change type, transparent concern tags and priority rationale, official source links, and a stable deduplication key.

Built for

• fleet operations and fleet risk monitoring teams prioritizing service-related document review
• warranty, service, and dealership teams tracking manufacturer guidance
• automotive compliance and product-quality analysts reviewing new or revised communications
• data teams feeding traceable NHTSA document changes into internal workflows

Why this instead of a raw TSB scraper

• Monitor changes, not just rows: persistent snapshots identify initial, new, and revised communications and suppress unchanged results by default.
• Focus the review queue: filter by make, model, model year, keyword, concern category, and minimum priority before results are emitted.
• See why an item was ranked: every result carries exact evidence terms and deterministic priority rationale rather than an unexplained score.
• Keep results traceable: retain the verbatim agency summary, official archive and dataset URLs, detection time, and stable dedupeKey.
• Process large source files predictably: the Actor streams and bounds the official ZIP/CSV workflow instead of loading the full archive into memory.

This is not a raw TSB export, defect detector, recall predictor, or legal/safety determination. A communication can be routine service guidance. priority only ranks explicit words in NHTSA's concise summary so a human can decide what to review.

What it emits

One paid manufacturer-communication-signal per matching document + make + model + model-year identity:

• signalType: initial, new, revised, or optionally unchanged
• NHTSA documentId, make, model, modelYears, and verbatim summary
• transparent concernCategories and exact evidenceTerms
• deterministic priority (low, medium, high, critical) and priorityRationale
• changedFields (summary for a revised stable identity)
• official archive and NHTSA dataset URLs, detection time, and stable dedupeKey

critical is reserved for explicit severe phrases such as “vehicle fire,” “loss of braking,” “loss of steering,” “air bag may not deploy,” or “loss of motive power.” It means “review first,” never “confirmed defect.”

Example input

{
  "makes": ["FORD"],
  "modelYears": ["2025"],
  "keywords": ["brake", "fire"],
  "concernCategories": ["braking", "fire-thermal"],
  "minimumPriority": "high",
  "includeInitial": true,
  "emitUnchanged": false,
  "maxSourceRecords": 200000,
  "maxItems": 100,
  "requestTimeoutSecs": 60,
  "baselineStoreName": "my-nhtsa-comms-baseline"
}

All make/model/year filters are exact and case-insensitive. Keyword filters are case-insensitive summary substrings. Multiple keywords and concern categories use OR within their group; different filter groups combine with AND.

Delta semantics

• initial — matching identity first seen while the named baseline is fresh
• new — previously unseen identity after that baseline has been initialized
• revised — same stable identity but normalized summary fingerprint changed
• unchanged — same identity and summary fingerprint; suppressed unless emitUnchanged is true

includeInitial: false builds the baseline without charging initial records. Use a dedicated baselineStoreName per monitoring configuration when you want independent history. Filtering happens before baseline access and before PPE charging, so materially changing filters is best paired with a new store name.

Each snapshot and its per-record pending-delivery state are persisted before its PPE dataset push, and a successful push is acknowledged afterward. This provides at-least-once delivery, not exactly-once delivery: if the run stops after a successful push but before its acknowledgement is persisted, the next run retries that signal and can produce a duplicate. Once A is acknowledged, a later failure on B retries B without pushing A again. Consumers that maintain current state should upsert/idempotently process records by the stable dedupeKey; append-only consumers should use it with the signal's business fields to detect retry duplicates while still retaining genuine later revisions. Snapshots use SHA-256-derived legal KV keys with 256 two-hex partition prefixes (record-00-… through record-ff-…) rather than one oversized JSON value. If a run stops because maxItems or the platform event-charge limit is reached, later source rows are intentionally not consumed or baselined.

Resource and source boundaries

Default source: MFR_COMMS_RECEIVED_2025-2026.zip, currently one CSV with these required columns:

TSB/Document ID, Make, Model, Model Year, Concise Summary

The downloader accepts HTTPS only, follows redirects, enforces a 25 MB compressed-download limit and timeout through body consumption, validates ZIP magic across arbitrary network chunks, and writes to a temporary file. unzip -p streams decompression through an RFC 4180 parser; neither the ~80 MB CSV text nor all records are retained. Uncompressed data is capped at 160 MB and any single CSV record at 2 MB measured in actual UTF-8 input bytes, including multiline content and record syntax. In-run duplicate tracking retains only fixed-size SHA-256 digests and is capped by maxSourceRecords, which also bounds consumed rows. Override sourceUrl only with an official archive that follows the same format.

Concern rules

Rules are deterministic case-insensitive phrase/token matches and may assign several categories:

• fire-thermal: fire, thermal, overheat, smoke, burning, melting terms
• braking: brake/braking/ABS terms
• steering: steering system terms
• airbag-restraint: airbag, seat-belt, restraint terms
• engine-stall-damage: stall, motive-power loss, engine damage/failure terms
• electrical: short circuit, wiring, battery, charging, fuse, voltage terms
• visibility: windshield, camera image, wiper, defroster terms
• software: software, firmware, programming, calibration-update terms
• warranty-service: warranty/service-campaign/inspection terms (inspection alone remains low unless another component or severity rule also matches)
• other: no category rule matched

Priorities are then assigned in order: explicit severe phrase → critical; failure/loss/smoke/crash/injury/stall language → high; a non-software/non-service component category → medium; otherwise → low. The exact matched terms and rationale are included in every result.

Local development

Requires Bun and the unzip executable.

bun install --frozen-lockfile
bun run test
bun run typecheck
npx --yes apify-cli validate-schema
APIFY_LOCAL_STORAGE_DIR=./storage bun run start

For local input, write storage/key_value_stores/default/INPUT.json. Reuse the same local storage and baseline name to test unchanged suppression.

Responsible use

The archive and summaries are official public NHTSA material, but users should verify every signal against the underlying communication and current NHTSA recall/vehicle records. Text matching can miss synonyms or flag routine guidance. Do not use this output alone for safety decisions, legal claims, recall status, automated adverse action, or claims that a manufacturer concealed or confirmed a defect.