Veritas — Hallucination-proof research

Frontend (React + Vite) and backend (Python research pipeline) for literature review with traceable quotes: every claim is cited, every quote is verified against the source PDF, and every conclusion is verifiable. Prize Winner @ Tartan Hacks.

---

App Workflow

1. Initial input

Enter your research question in the text area and optionally add PDFs by dragging and dropping or clicking to browse. Trust badges remind you that every claim is cited, every quote is traceable, and every conclusion is verifiable. Press Enter to submit or Shift+Enter for a new line.

2. Processing

While the pipeline runs, a step-by-step progress view shows where you are: finding sources → extracting quotes → cross-checking claims → compiling review. The active step shows live sub-tasks (e.g. which chunk is being analyzed), and a progress bar reflects overall completion.

3. Written results

After the run completes, the Literature Review section shows counts for sources, quotes, and findings. You can download the Comprehensive Research Report (full literature review) or the Executive Summary as PDF. The summary gives a concise, evidence-based overview of the research question.

4. Sources

Each source appears as an expandable card. Expand one to see its title, verification status, and all extracted quotes. Every quote is verbatim from the paper and traceable to the source, so you can verify claims yourself.

---

Architecture overview

Layer	Role
Frontend (tartan_frontend/)	React app: research query input, optional PDF upload, progress stepper, results (summary + full literature review + source cards, PDF download).
API (app.py)	FastAPI: POST /api/research runs the pipeline, streams NDJSON progress and final result; GET /api/papers/<name> serves source PDFs.
Pipeline (tartan_backend/)	run_all.py → research_bot (extract quotes) → clean_quotes_in_place → merge_quote_csvs → synthesize_ideas → summarize_review + generate_literature_review.
Sources	With no PDFs: query → arXiv + Semantic Scholar + OpenAlex → LLM rank → download top PDFs. With PDFs: use uploads as sources.

Key backend modules are documented at the top of each file (usage, environment, and pipeline role).

---

Run the full stack

1. Backend (API + pipeline)
   From the project root (tartan26/), with a virtualenv that has the backend deps and DEDALUS_API_KEY set (e.g. in tartan_backend/.env):

   cd /path/to/tartan26
   pip install -r requirements.txt
   uvicorn app:app --reload --port 8000

   (The API lives in the root; running from tartan_backend/ will fail with "Could not import module app".)

2. Frontend
   In another terminal:

   cd tartan_frontend && npm install && npm run dev

3. Open http://localhost:5173 — enter a research query, optionally upload PDFs, and submit. The app proxies /api to the backend and shows sources, executive summary, and full literature review when the pipeline finishes.

---

Backend-only (CLI pipeline)

From tartan_backend/:

python run_all.py --rq "Your research question" --papers_dir ./papers --with_ideas

Put PDFs in papers/ and set DEDALUS_API_KEY in tartan_backend/.env. See docstrings in run_all.py, research_bot.py, and related scripts for options and pipeline order.

---

Documentation

• In-code: Every backend Python module and frontend TS/TSX file has a file-level docstring (or block comment) describing purpose, usage, and environment.
• Model choices: MODEL_ARCHITECTURE.md explains which LLM is used at each pipeline stage and why.
• MCP: MCP_SERVER_README.md describes the Semantic Scholar MCP server and how to run it.