# DeepDoc

A local deep-research tool that turns a folder of PDF, Word, PowerPoint, image and text files into a structured Markdown report, using a judgment model to check evidence coverage alongside the generative LLM.

DeepDoc is a local deep-research tool that turns a directory of PDF, DOCX, PPTX, image, TXT or Markdown files into one structured Markdown report. It extracts and indexes the content into Qdrant, plans a report structure from the topic and outline, which the user can approve or revise, then runs one research agent per report section.

Each section's agent generates research queries, retrieves candidate chunks from Qdrant, and uses TypeSafe's Jev model to judge each chunk's relevance, usable evidence, contradiction and prompt-injection-like content before it reaches the writing LLM. Jev's reflection scores also flag subsections that still lack evidence, triggering new queries up to a configured retry limit; the configured LLM still handles planning, query generation, synthesis and the actual writing.

The final report writer combines all completed sections into one exported Markdown file. Filtering thresholds, ranking, retry limits and routing are ordinary Python logic built on Jev's probability outputs.

Website: https://github.com/Oqura-ai/deepdoc