kazol196295/medical_report_simplifier

1. Purpose and Entry Point

This is a session-oriented Streamlit application that converts uploaded medical-report images into OCR text, AI explanations, report-grounded chat answers, and optional clinical-trial matches. The executable UI entry point is app.py; it coordinates all processing and stores transient interaction state in the Streamlit session.

Paths: app.py, .streamlit/config.toml, .streamlit/secrets.toml.example

2. Upload and OCR Boundary

Image uploads (JPG/PNG/JPEG/BMP) cross into the local OCR layer, which uses the host-installed Tesseract engine to produce raw report text. OCR is a material local-system dependency and the first transformation boundary: image bytes become text usable by every downstream feature.

Paths: app.py, src/ocr_tesseract.py, packages.txt

3. Report Analysis Agent

The conversational/analysis layer is modeled as a LangGraph state graph. It accepts OCR text, requested analysis mode, and conversation context; selects or invokes analysis tools; and generates explanatory medical content through Groq-hosted LLaMA 3.3-70B. Prompts and structured output expectations are separated from graph control flow.

Paths: src/langgraph_agent.py, src/tools.py, src/prompts.py

4. RAG Index and Report-Grounded Chat

Extracted text is indexed into a FAISS-backed retrieval layer using Hugging Face/Sentence Transformer embeddings. Follow-up chat retrieves semantically relevant report fragments before LLM response generation, creating a grounding boundary between report-specific evidence and the external model. The index is intended to be session-scoped rather than persistent user storage.

Paths: src/rag_engine.py, src/langgraph_agent.py, app.py

5. Patient Profile Extraction

For trial matching, OCR text is transformed into an editable structured patient profile: age, gender, diagnosis, biomarkers, and conditions. This is a separate AI extraction stage so users can review or correct uncertain OCR/LLM-derived fields before they drive external trial queries.

Paths: src/patient_profiler.py, src/prompts.py, app.py

6. Clinical-Trial Search and Matching

The approved profile is sent to the ClinicalTrials.gov API v2 client to find recruiting trials. Trial results are then evaluated for eligibility and presented with matched/unmet criteria. This stage crosses an external public-data boundary; it should tolerate unavailable or sparse API results independently of analysis/chat functionality.

Paths: src/clinical_trials_api.py, src/patient_profiler.py, src/tools.py

7. Runtime and External Dependencies

Runtime is Python 3.9+ with Streamlit as the UI server. Required external services are Groq (API key from Streamlit secrets, environment, or user entry) and ClinicalTrials.gov. Local infrastructure includes Tesseract plus native libraries, FAISS vector search, and model/embedding dependencies. Deployment is compatible with Streamlit Cloud, with OS packages declared separately.

Paths: requirements.txt, runtime.txt, packages.txt