Grounded question answering
The model receives selected source passages instead of the entire document and is instructed to answer from those passages.
Early-stage open-source project
Ask your documents.
See the evidence.
DocAI is a document-intelligence assistant for PDF and TXT files. It retrieves relevant passages, generates concise answers from those sources, and stores citations so you can inspect the evidence behind every response.
PDF + TXT
Grounded answers
Stored citations
research-notes.pdf
18 pages · processed
You
What are the main limitations described in this document?
AI
Grounded answer
The document identifies three main limitations: restricted input coverage, dependence on text quality, and the need for human review in high-stakes settings. [1] [2]
1
research-notes.pdf · passage 12
2
research-notes.pdf · passage 17
7 MB
upload limit
PDF / TXT
current file support
User-scoped
document access
Inspectable
stored citations
Built for evidence-aware Q&A
Useful answers without hiding the source.
DocAI keeps the retrieval path understandable: uploaded documents are processed into searchable passages, relevant passages are selected, and the answer is generated from those sources.
Persistent citations
Answers keep references to the source document, passage, filename, and excerpt so the evidence trail remains available for review.
Transparent retrieval
Current retrieval is deliberately lexical and explainable, using direct terms, related terms, title matches, and exact-phrase bonuses.
User-scoped data
Protected procedures check the signed-in user before accessing documents, conversations, messages, and citations.
Controlled ingestion
PDF and TXT uploads are validated by extension, MIME type, file size, and readable text before processing.
Full-stack foundation
React, TypeScript, tRPC, Express, MySQL/Drizzle, object storage, and automated tests form the current implementation.
How it works
From upload to cited answer.
DocAI is designed to keep the path between a document and an answer explicit.
-
1
Upload a document
An authenticated user uploads a text-based PDF or TXT file. Unsupported or unreadable files are rejected.
-
2
Extract and chunk text
Readable text is normalized and divided into overlapping chunks for search and retrieval.
-
3
Retrieve relevant passages
The current ranking system scores lexical signals to select passages related to the user's question.
-
4
Generate from the sources
Only the selected passages are sent to the configured language-model integration with grounding instructions.
-
5
Save the evidence trail
The response, conversation history, and citations are persisted so users can return to the answer and inspect its sources.
Project principles
Grounded by design. Clear about limitations.
Evidence over opacity
A generated answer should make it easier—not harder—to inspect the material it relies on.
Explicit boundaries
A citation does not make an answer automatically correct. Users should review original documents and cited excerpts.
Practical iteration
DocAI is in active development. Current limitations include no OCR for scanned image-only PDFs and lexical rather than embedding-based retrieval.
Open source · In development
Explore the implementation.
Read the code, architecture notes, tests, and current limitations in the public repository.