Extract text from a PDF or read a plain text file, then create overlapping chunks.
SOFTWARE PROJECT · 2026
A small RAG system for asking questions about your own documents
I built DocuMindGPT as a command line workflow for PDFs and text files. It stores document chunks in Supabase, retrieves relevant context for each question, and asks a second model to audit the generated answer.
- Input
- PDF and text
- Storage
- Supabase pgvector
- Generator
- Gemini 2.5 Flash
- Interface
- Python CLI
PURPOSE
Keep answers tied to a supplied source
The tool is meant for situations where the answer should come from a known document rather than the model memory. A system instruction tells the generator to use only the retrieved context and to say when the context does not contain an answer.
I separated ingestion from chat so a document can be processed once and queried repeatedly. The command line keeps the example compact and makes each stage easy to inspect.
WORKFLOW
From a file to an audited answer
Generate Gemini document embeddings and insert the chunks into Supabase.
Embed the question, retrieve matching chunks, and ask Gemini to answer from that context.
Send the question, context, and answer to a second evaluation prompt.
Chunk size1,000 characters
Overlap200 characters
RetrievedTop 5 chunks
Threshold0.5 similarity
ANSWER AUDIT
A useful check, not an independent judge
The answer model is instructed not to add information beyond the supplied chunks.
The evaluator checks grounding and relevance, then returns structured JSON.
The score runs from 1 through 10. A score of 7 or more receives a Pass verdict. If retrieval returns no context, the audit is skipped.
The generator and evaluator both use Gemini 2.5 Flash. The audit can help during interactive use, but it is still a model based assessment and should not be treated as independent verification.
CURRENT EVIDENCE
What is public today
The repository does not publish aggregate benchmark results
The code implements retrieval, grounded generation, and a per answer audit. It does not currently include a benchmark dataset, a saved evaluation run, or an aggregate accuracy table.
- Implemented
PDF and text ingestion, chunking, batch embedding, Supabase insertion, similarity retrieval, grounded prompting, interactive chat, and JSON audit parsing.
- Handled failures
The CLI reports missing files, empty PDFs, missing credentials, failed API calls, absent retrieval context, and malformed evaluator output.
- Still needed for comparative claims
A fixed evaluation set, retrieval measures, answer quality measures, and comparison with a baseline under the same documents and questions.
SETUP
Run the command line tool
Create a Supabase project, run `setup.sql`, and place the Gemini and Supabase credentials in a local `.env` file. Credentials should not be committed.
pip install -r requirements.txt
python main.py upload path/to/document.pdf
python main.py chat