UWaterlooRFRChatBot / README.md
Anna Zhou
Use OAuth for Drive request uploads
2f024ae
|
Raw History Blame Contribute Delete
2.76 kB
---
title: RFR Knowledge Platform
emoji: 🌊
colorFrom: blue
colorTo: green
sdk: streamlit
app_file: app/app.py
pinned: false
---
# RFR Co-op: LLM-Assisted Managed Retreat RAG Pipeline
A RAG (Retrieval-Augmented Generation) pipeline for extracting and synthesizing insights from managed retreat case study literature, built for the **Retreat From Risk (RFR)** research initiative.
## Setup
```bash
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
```
## Usage
```bash
# 1. Set API keys
export ANTHROPIC_API_KEY=your_key_here
export OPENAI_API_KEY=your_key_here
# 2. Convert canonical Ver2 memos to JSON chunks, then build embeddings
python ingest/ingest_pipeline_outputs.py
python ingest/create_summary_chunks.py
python ingest/embed_and_index.py
# 3. Run CLI pipeline
python rag/pipeline.py
# 4. Launch Streamlit web app
streamlit run app/app.py
```
## Hugging Face request intake
The **New Case Study Requests** tab saves each submitted request to Google Drive,
under folder `10n_-uOCT2GXu_G3r1qi5qJr8hHoy-E_y`. In Hugging Face Spaces, use
OAuth credentials for the Google account that should own submitted files, usually
`rfr.rag.uwaterloo@gmail.com`. Set these secrets:
```text
GOOGLE_OAUTH_CLIENT_ID
GOOGLE_OAUTH_CLIENT_SECRET
GOOGLE_OAUTH_REFRESH_TOKEN
```
To generate those values, run:
```bash
python3 -m case_study_pipeline.create_drive_oauth_secrets \
--credentials-file /path/to/oauth-client.json \
--space-id UWRFR/UWaterlooRFRChatBot
```
To use a different parent folder, set `RFR_REQUESTS_DRIVE_FOLDER_ID`.
## Ingesting documents
```bash
# Convert canonical case-study memos into section chunks
python ingest/ingest_pipeline_outputs.py
# Refresh overview-summary chunks for broad comparison queries
python ingest/create_summary_chunks.py
# Build the precomputed embedding store used by rag/retriever.py
python ingest/embed_and_index.py
```
## Project structure
```
rfr-rag/
β”œβ”€β”€ data/raw/ # Original PDFs and spreadsheets (not committed)
β”œβ”€β”€ data/extracted/ # Ver2-derived JSON chunks + embedding store
β”œβ”€β”€ ingest/ # Ver2 ingestion, summary chunks, embedding script
β”œβ”€β”€ case_study_pipeline/# Source extraction/audit workflow
β”œβ”€β”€ rag/ # Retriever, prompt builder, pipeline
β”œβ”€β”€ persona/ # Decision-maker persona config
β”œβ”€β”€ questionnaire/ # Guided questionnaire structure
β”œβ”€β”€ app/ # Streamlit interface
β”œβ”€β”€ docs/ # Extraction schema, memos
└── tests/
```
See [CLAUDE.md](CLAUDE.md) for full project context.
For day-to-day website upkeep and adding new case studies, see
[Website Maintenance and New Case Study Guide](docs/website_maintenance.md).