# 🚀 RAG Backend Setup & Testing Guide
## ✅ Current Status
**Server is running!** ✅
- Health check: ✅ Working
- Document upload: ✅ Working
- Vector store: ✅ Working (ChromaDB)
- Retrieval: ✅ Working (0.467 confidence on test query)
- LLM generation: ⚠️ Needs API key
## 📝 Step 1: Add Your Gemini API Key
1. **Get your API key** (if you don't have one):
- Go to: https://aistudio.google.com/app/apikey
- Create a new API key (it's FREE)
2. **Add to `.env` file**:
```powershell
# Open .env file in rag-backend folder
notepad .env
```
3. **Set the key**:
```env
GEMINI_API_KEY=your_actual_api_key_here
LLM_PROVIDER=gemini
```
4. **Restart the server** (if it's running):
- Stop current server (Ctrl+C)
- Start again: `.\venv\Scripts\python.exe -m uvicorn app.main:app --reload --port 8000`
## 🧪 Step 2: Test with Your Real Documents
### Option A: Using Swagger UI (Easiest)
1. Open: **http://localhost:8000/docs**
2. Find `POST /kb/upload`
3. Click "Try it out"
4. Upload your PDF/DOCX file
5. Set:
- `user_id`: your user ID
- `kb_id`: your KB ID
6. Click "Execute"
7. Wait 5-10 seconds for processing
### Option B: Using curl
```powershell
curl -X POST http://localhost:8000/kb/upload `
-F "file=@your_document.pdf" `
-F "user_id=user123" `
-F "kb_id=kb001"
```
## 🔍 Step 3: Check Retrieval Quality
**BEFORE testing chat**, verify retrieval works:
```powershell
# Test search endpoint
curl -X POST "http://localhost:8000/kb/search?query=your%20question&kb_id=kb001&user_id=user123&top_k=5"
```
**What to look for:**
- ✅ Similarity scores > 0.35 (good)
- ✅ Relevant chunks returned
- ✅ Correct file names in metadata
**If scores are low (< 0.3):**
- Try rephrasing your question
- Check if document was fully processed
- Verify chunks were created (check `/kb/stats`)
## 💬 Step 4: Test Full Chat
Once retrieval looks good, test chat:
```powershell
curl -X POST http://localhost:8000/chat `
-H "Content-Type: application/json" `
-d '{\"user_id\":\"user123\",\"kb_id\":\"kb001\",\"question\":\"Your question here\"}'
```
**Expected response:**
- ✅ Answer based on your documents
- ✅ Citations with file names
- ✅ Confidence score > 0.35
- ✅ `from_knowledge_base: true`
## 🐛 Troubleshooting
### Issue: "LLM not configured"
**Fix:** Add `GEMINI_API_KEY` to `.env` file
### Issue: Low similarity scores
**Fix:**
- Adjust threshold: Set `SIMILARITY_THRESHOLD=0.25` in `.env` (lower = more lenient)
- Re-upload document with better formatting
- Try different query phrasing
### Issue: "No chunks found"
**Fix:**
- Wait 10-15 seconds after upload
- Check `/kb/stats` to verify chunks exist
- Check server logs for parsing errors
### Issue: Hallucination (wrong answers)
**Fix:**
- Increase `SIMILARITY_THRESHOLD=0.45` (stricter)
- Check retrieval quality first (Step 3)
- Verify citations are present
## 📊 Quick Test Script
Run the automated test:
```powershell
.\venv\Scripts\python.exe test_rag.py
```
## 🎯 Next Steps
1. ✅ Add Gemini API key
2. ✅ Upload your real KB documents
3. ✅ Test retrieval quality
4. ✅ Test chat with real questions
5. ✅ Adjust thresholds based on results
## 📞 Need Help?
Check the logs in the terminal where the server is running for detailed error messages.
