Data processing · updated 2026-09-15
How a document becomes a reviewable comparison.
The V0.1 pipeline separates reading, candidate proposal, deterministic checks, human review, and report export. Each stage keeps its result so a partial or uncertain run remains visible.
Pipeline stages
- Upload validation. Check extension, file type, size, and safety limits.
- Parsing. Read text, tables, and locations from PDF, DOCX, XLSX, CSV, or offline HTML.
- Candidate proposal. A configured model or the offline demo extractor suggests products, fields, conditions, and values.
- Evidence checks. Confirm that quotes belong to the source block, values can be normalized, and product ownership is plausible.
- Comparison and review. Group observations by product, field, and condition; keep Missing, Conflict, and Unsupported states explicit.
- Snapshot and export. Freeze one reviewed revision and render CSV, Markdown, JSON, or PDF from that snapshot.
What is sent to a model
When a model is configured, the service sends only allowed document blocks and the fixed extraction contract to the model provider configured by the operator. The model proposes candidates; it does not get tools, browse the web, execute document instructions, or decide that a value is Supported by itself.
Deterministic checks run after the model response. Invalid JSON, missing quotes, wrong units, cross-product values, and unsupported ranges remain visible for review.
What is stored for evidence
| Category | Purpose | Default retention |
|---|---|---|
| Uploaded files | Parse the requested comparison and preserve the original source. | 7 days |
| Document blocks | Keep text, tables, and page/sheet/record locators for evidence. | 7 days |
| Observations and cells | Build the canonical comparison and show uncertainty. | 7 days |
| Report snapshot | Render consistent exports after review. | 7 days |
| Analytics | Measure anonymous product funnel events. | Provider settings |
Hard boundaries
- No OCR for scanned pages in V0.1.
- No URL crawling, spreadsheet formula execution, macros, or external links.
- No guessed values when a field is absent.
- No independent testing or certification of manufacturer claims.
- No public indexing of private comparison routes; user documents stay behind session access.
中文说明
流程分为上传校验、解析、候选提议、证据核验、人工复核和报告快照。模型只提出候选，固定规则核验单位、引用、产品归属和冲突。没有模型时使用演示抽取器。
V0.1 不做扫描件 OCR、网址抓取、公式或宏执行，不猜测缺失值，也不验证厂商声明。文件、证据和报告默认保留 7 天。
Questions and corrections
For a correction or a data-processing question, contact llzclm@gmail.com. Please describe the issue without sending confidential document contents.