Document Reader
System / Product: Litigant Portal
Model Type: Prompted language model for documents, images, or text
Version: 1.0
Date: 2026-11-18
Owner: Free Law Project, individual owner pending confirmation
Status: Draft. Live deployment and model selection await confirmation.
1. Purpose
1.1 What this model does
A litigant can attach a document or image and ask a question such as which date a notice gives for a hearing. Small, supported attachments can go directly to the chat assistant. For an attachment that needs a separate read, the assistant calls QueryDocument with the upload ID and a specific request. The reader returns text that the assistant can use in its reply. Attachment handling · Document reader
1.2 Intended users
The assistant receives the answer as a tool result. The litigant can expand the document card to see the request and response. If the assistant uses the result for a guided form, the litigant can inspect and confirm the proposed value. Reader result card · Fact review card
1.3 Scope and limitations
The reader answers from one file per call. The tool looks up the upload against the identity that owns the chat thread. It cannot establish that a document is authentic, that a deadline is legally effective, or that other documents in the case agree. Upload lookup and request
2. Base Model
| Field | Details |
|---|---|
| Base model name | The reader prefers the site’s assistant model when it can read the input. It otherwise selects the first enabled model with the required capability in curated order. |
| Version / checkpoint | May vary by site, file type, and catalog release. The deployed releases are unverified. |
| Provider | Amazon Bedrock. The model developer depends on the selected family. |
| License | Model-specific terms. Confirm each deployed reader family and its applicable terms. |
| Model location | Public reference for the deployed model awaits confirmation. Reader selection · Bedrock routing |
2.1 Model Modifications
Prompting only. FLP uses provider-hosted weights and controls the instructions sent with each request.
3. Data
3.1 Pretraining Dataset
FLP used the hosted model as supplied. No FLP pretraining dataset applies.
3.2 Finetuning Dataset
No FLP finetuning dataset applies.
3.3 Prompt Design and Versioning
The versioned reader prompt asks for a concise factual answer, exact quotes where useful, and an explicit acknowledgement if the file lacks the requested information. The assistant is separately instructed to request a date and its supporting quote when collecting a flow fact and to ask again if the needed date was omitted. Audit reconstruction uses the tool source at the recorded code revision. Reader prompt · Assistant instructions · Audit guide
3.4 Validation and Test Dataset
Repository fixtures exercise text, PDF, image, and office-document handling, oversized inputs, missing uploads, and reader selection. The routing and truncation tests use controlled model responses. No labeled collection of representative litigant uploads with verified extractions was identified for this draft. Upload tests · Reader model tests
3.5 Label Documentation
No FLP training labels apply.
4. Training & Evaluation
4.1 Design Decisions and Rationale
Inlining a supported file lets the assistant answer directly. Size, format, model capabilities, and the remaining request budget determine when it receives an attachment note instead. A separate reader call can target a particular fact without resending the rest of the chat. The upload limit is 20 MiB. The reader checks PDFs with a known page count against a 100-page cap and checks extracted text for text-kind files against a 150,000-token cap. File limits and dispatch
4.2 Metrics and Evaluation
The available tests verify that the selected route can accept the content type and that truncated output is marked incomplete. They supply no estimate of date extraction accuracy, quotation fidelity, or performance on scans. The chat benchmark does not grade this reader separately. Reader tests · Benchmark cases
4.3 Failure Analysis
A missing, unreadable, or oversized file can yield an explanatory error for the assistant and a generic failure on the result card. If generation reaches its output limit, the reader prefixes the partial answer with an incompleteness notice. A fluent answer that silently omits a date or misquotes a sentence can still reach the user. Error and truncation handling · Error display
5. Known Limitations
Legibility, layout, handwriting, language, and similar dates on a single notice can change which text the model selects. The reviewed evidence contains no accuracy estimate by format or user group. A model-supplied quote is useful for locating a passage, but the reader does not mechanically verify that every quote matches the file.
6. Deployment and Monitoring
6.1 Deployment Setup and Data Dependencies
Django retrieves the selected stored file and sends the relevant text, image, or document content, filename, and question to Bedrock. The tool returns the response to chat and supplies the filename, request, and answer for the expandable card. A later assistant message may repeat or paraphrase the result. Tool implementation · Chat integration
6.2 Monitoring Plan
Available signals include reader truncation logs, tool errors, usage cost when reported, and the saved request and result in a thread audit. These identify operational issues. Measuring silent extraction mistakes requires comparing results with source files. Production review ownership, thresholds, and cadence await confirmation. Reader diagnostics · Audit workflow
6.3 Re-evaluation Criteria and Process
Proposed triggers are a new reader family, changes to file formats or limits, or a reported incorrect extraction. Compare a permitted sample with verified source passages across file types, including a file with multiple dates and one without the requested fact.
7. Version History
| Card version | Date | Author | Change |
|---|---|---|---|
| 0.1 | 2026-11-18 | Codex, prepared for FLP review | Initial draft against portal revision b20fec4. |
8. Output Disclosure
The expanded tool card shows “Searched” with the filename, then “Request” and “Response.” Those labels identify the file and action. They do not identify the response as a model interpretation or establish that quoted text was verified. The inline path has only the ordinary chat reply. A proposed improvement is to label the reader response as AI-generated beside the text and preserve that attribution when extracted material is reused. Document result display · Litigant assistant
9. Contacts
| Role | Name | Contact |
|---|---|---|
| Feature owner | Litigant Portal team. | Public contact pending confirmation. |
| AI team contact | John Gardner | john@free.law |