Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/sputhenofficial/claimjumper/llms.txt

Use this file to discover all available pages before exploring further.

ClaimJumper processes a PNG remittance image through five sequential steps. Each step has a clear input/output contract and can be tested independently using injected dependencies. Model calls are isolated to ingest, triage, and draft — prioritization and invariant enforcement are deterministic code that never touch the OpenAI API.

​
Pipeline diagram

The following diagram shows the full flow from image upload to human-approved download: 
PNG EOB/remittance
  → structured ingest
  → denial triage + invariant enforcement
  → deterministic prioritization
  → draft or human-review brief
  → human approval
  → citation-preserving download

​
Modules

​
Ingest

lib/pipeline/ingest.ts sends the PNG to GPT-5.6 Sol via the OpenAI Responses API using a base64-encoded image input. The model extracts every claim, service line, and adjustment code legend entry into a strict IngestOutput validated by Zod. Results are cached to disk using a SHA-256 + schema-version key so the same image never triggers a redundant API call during development or testing. See the Ingest page for the full function signature, caching strategy, output schema, and prompt design.

​
Triage

lib/pipeline/triage.ts sends each AdjustmentLine to GPT-5.6 Sol for a lane decision. The model reads Group Code, CARC, and RARC together and returns a structured TriageModelOutput containing the lane, root cause, rationale, evidence needed, confidence, and recovery probability. See the Triage page for all model output fields, prompt constraints, and invariant integration.

​
Invariants

lib/pipeline/invariants.ts runs synchronously after every triage model call — before the result is returned to the caller. It enforces three non-negotiable safety rules in code:
  • A CO group-code adjustment can never be routed to patient_bill.
  • A CARC 197 denial can never be automatically written off.
  • Any triage with confidence < 0.6 is escalated to human_review.
When any rule fires, the lane is overridden to human_review and the override reasons are attached as a string[] on the DenialTriage object. Invariant enforcement is covered inline on the Triage page.

​
Prioritize

lib/pipeline/prioritize.ts ranks triaged denial lines by appeal deadline, urgency, and a weighted financial score. Every calculation is deterministic arithmetic — no model calls occur in this step. Lines sort first by deadline ascending, then by priority_score descending within the same deadline. See the Prioritize page for the score formula, deadline arithmetic, urgency calculation, and exported helper functions.

​
Draft

lib/pipeline/draft.ts produces the work artifact for each line in the queue. For appeal and corrected_claim lanes it calls GPT-5.6 Terra, then validates every citation against the original parsed data before the artifact is returned. For human_review it constructs a deterministic plain-text brief without a model call. For patient_bill and write_off it throws — those lanes never receive drafted artifacts. See the Draft page for citation validation, content safety checks, and prompt design.

​
Dependency injection

All pipeline functions accept an optional dependencies argument so tests can substitute mock implementations without live API calls or a real filesystem. The runTriagePipeline() function in app/api/triage/route.ts is the server-side composition point — it wires each module together and exposes the same seam for integration tests: 
interface PipelineDependencies {
  ingest?: (imageBytes: Uint8Array) => Promise<IngestOutput>;
  triage?: typeof triage;
  draft?: DraftModelClient;
  today?: () => string;
}

export async function runTriagePipeline(
  imageBytes: Uint8Array,
  dependencies: PipelineDependencies = {},
): Promise<TriageResponse>
Injected triage functions must return an already-enforced DenialTriage — the same guarantee that triage() itself provides by always calling enforceInvariants() before returning.

​
Model configuration

All model IDs are declared in a single file so they can be updated in one place: 
export const MODELS = {
  ingest: "gpt-5.6-sol",
  triage: "gpt-5.6-sol",
  draft: "gpt-5.6-terra",
} as const;
MODELS.ingest and MODELS.triage use GPT-5.6 Sol for its vision capability and structured-output reliability. MODELS.draft uses GPT-5.6 Terra for its grounded, citation-quality text generation.
All OpenAI calls in ClaimJumper go through lib/openai.ts — the fetch() API is never called directly. responseWithSchema() handles text input and responseWithImageSchema() handles base64-encoded PNG input; both enforce Zod-validated structured outputs and read OPENAI_API_KEY from the environment.

Build docs developers (and LLMs) love

Get started for freeTalk to us