# SpecSheet AI

> SpecSheet AI turns supplier documents into a source-backed technical product comparison that people can review.

## Product definition

SpecSheet AI is a focused comparison tool for procurement, presales, and engineering teams. It reads PDF, DOCX, XLSX, CSV, and offline HTML files, identifies product and specification candidates, maps them to a canonical field set, and keeps the source location with each value.

## What the MVP does

- Reads text, tables, and locations from uploaded documents.
- Identifies products, fields, conditions, and value candidates.
- Normalizes a limited allow-list of technical units with Decimal arithmetic.
- Compares products using the states Supported, Missing, Conflict, and Unsupported.
- Preserves source, evidence quote, and document location for review.
- Exports the reviewed snapshot as CSV, Markdown, JSON, or PDF.

## How uncertainty is handled

- **Supported** means a value was found with a checked source reference. It is not independent verification of the technical claim.
- **Missing** means no candidate was found in sources that were successfully read and correctly associated with the product.
- **Conflict** means incompatible candidates remain for the same product, field, and condition.
- **Unsupported** means a candidate or its evidence cannot be verified from the source.

## Analysis boundary

The parser reads the documents first. When configured, an OpenAI-compatible model proposes candidates; deterministic rules then check units, values, product ownership, evidence, and conflicts. Without a model key, the hosted trial uses its configured demo extractor. The MVP does not run OCR, fetch URLs, execute spreadsheet formulas or macros, guess missing values, or verify manufacturer claims.

## Core fields

The initial dictionary includes net mass, gross mass, dimensions, rated input power, supply voltage, supply frequency, warranty duration, and construction material. Extensions use the `coffee.*`, `motorcycle.*`, `industrial.*`, and `custom.*` namespaces without pretending to be an industry database.

## Try it

- Hosted MVP: https://specsheet-ai.pages.dev/
- Pricing and access: https://specsheet-ai.pages.dev/pricing.md
- Public source benchmark notes: https://github.com/llzclm1/specsheet-ai/blob/main/README.en.md

## Scope limits

This is V0.1. It is designed to make a supplier packet easier to compare and audit. It does not provide general document chat, OCR for scanned pages, URL crawling, payment automation, a PIM catalog, or a claim that extracted values are correct outside their cited source.