Architecture

Cell is organized into four systems: document extraction, application processing, agent prompts, and PDF operations. Each is independent — import only what you need.

Document extraction pipeline

The core of Cell is a multi-pass pipeline that turns insurance PDFs into structured, queryable data.

flowchart LR
  PDF[PDF Document] --> P0[Pass 0: Classification]
  P0 --> P1[Pass 1: Metadata]
  P1 --> P2[Pass 2: Sections]
  P2 --> P3[Pass 3: Enrichment]
  P3 --> OUT[Structured Data]

Pass 0 — Classification

Determines whether a document is a policy or a quote. Returns document type, confidence score, and supporting signals. Uses the classification model (fast/cheap).

Pass 1 — Metadata

Extracts high-level metadata: carrier, policy/quote number, dates, premium, insured name, and a coverage table with limits and deductibles. Uses the metadata model (capable).

Supports an onMetadata callback for early persistence — metadata is saved immediately so it survives downstream failures.

Pass 2 — Sections

Splits the document into page chunks (starting at 15 pages) and extracts structured sections with page-level provenance. Uses the sections model.

Adaptive fallback: if a chunk's output is truncated (JSON parse failure), Cell re-splits into smaller chunks (10 pages, then 5), and escalates to the sectionsFallback model with higher token limits. Results are merged across all chunks.

Pass 3 — Enrichment

A non-fatal pass that parses raw text into structured supplementary fields: regulatory context, complaint contacts, costs and fees, claims contacts. Uses the enrichment model. Failures here don't fail the pipeline.

Quote-specific extraction

Quotes run a variant pipeline (passes 1-2) that also extracts:

• Premium breakdowns — line-by-line premium details
• Subjectivities — conditions that must be met before binding
• Underwriting conditions — carrier requirements

Application processing

Cell provides prompt builders for the full application lifecycle:

flowchart LR
  D[Detection] --> F[Field Extraction]
  F --> A[Auto-Fill]
  A --> Q[Question Batching]
  Q --> P[Answer Parsing]
  P --> PDF[PDF Filling]

1. Detection — classify whether a PDF is an insurance application form
2. Field extraction — read every field as structured data (text, numeric, currency, date, yes/no, table, declaration)
3. Auto-fill — match fields against known business context
4. Question batching — organize unfilled fields into topic-based batches
5. Answer parsing — parse free-text replies into structured field values
6. PDF filling — write answers back onto the PDF (AcroForm or text overlay)

Agent prompt system

A composable system for building insurance-aware conversational agents:

buildAgentSystemPrompt(ctx)
  ├── Identity          — agent name, company context
  ├── Intent            — direct / mediated / observed behavior
  ├── Formatting        — platform-specific output rules
  ├── Safety            — scope guardrails, anti-hallucination
  ├── Coverage gaps     — gap detection guidance
  ├── COI routing       — certificate of insurance handling
  ├── Quotes/policies   — document type differentiation
  └── Memory            — cross-conversation continuity

Each module is independently importable for custom composition. The system supports five platforms (email, chat, SMS, Slack, Discord) and three communication intents (direct, mediated, observed).

Design principles

• Provider-agnostic — accepts any LanguageModel from the Vercel AI SDK. Default Anthropic models are lazy-loaded and optional.
• Pure TypeScript — no framework dependencies. Works in Node.js, Deno, edge runtimes.
• Fail gracefully — early persistence callbacks, non-fatal enrichment, adaptive chunk retry.
• Schema-only tools — tool definitions provide schemas without implementations, so consumers control execution.