fix: enforce strict JSON mode and resolve unicode crash in Groq client

reduce judge context size to prevent truncation

fix: add debug logging to judge agent

fix: remove load_dotenv from agents, use system env vars for HF compatibility

multi-model Groq fallback — llama-70b → llama-4-scout → gpt-oss-120b → llama-8b

fixed the loading key probelm

added model fallback avoidance

on demand pubmed lookup

wired the agents

added explainer agent plain English summaries with practical takeaways

added decomposer agent splits complex multi-part claims into sub-claims

added judge agent classification, stance detection, quality scoring

feat: Phase 2 complete — judge agent, evidence cards, transparency UI

judge agent in UI, evidence quality cards, study breakdown

fix: remove hardcoded paths for HuggingFace deployment

added working pipeline

added prompt for contradiction agent

correct file

setting up contradiction agent

feat: Phase 2 BioBERT embeddings, accuracy 80% → 90%, response time -38%

changed embeddings, used biobert for better similarity and better accuracy

expanded the database and batched embeddings

necessary changes for hf deploy

added location for load env

used json formatting for understandable output

fixed syntax

saved papers and retrieval logic

connected database and ingestion

added project structure basic