docs: update README and forge documentation

- README: Fixed backend command, added GPU compatibility reference
- forge.md: Fixed train CLI (--output-dir), added GPU troubleshooting
- Added reference to GPU compatibility guide for RTX 50-series

Dockerfile

@@ -28,10 +28,10 @@ COPY --from=builder /app/wheels /wheels
 RUN pip install --no-cache-dir /wheels/*
 
 # Copy application code
-COPY companion/ ./companion/
-COPY companion/forge/ ./companion/forge/
-COPY companion/indexer_daemon/ ./companion/indexer_daemon/
-COPY companion/rag/ ./companion/rag/
+COPY src/companion/ ./companion/
+COPY src/companion/forge/ ./companion/forge/
+COPY src/companion/indexer_daemon/ ./companion/indexer_daemon/
+COPY src/companion/rag/ ./companion/rag/
 
 # Create directories for data
 RUN mkdir -p /data/vectors /data/memory /models

Dockerfile.indexer

@@ -13,9 +13,9 @@ RUN pip install --no-cache-dir \
     pydantic lancedb pyarrow requests watchdog typer rich numpy httpx
 
 # Copy application code
-COPY companion/ ./companion/
-COPY companion/indexer_daemon/ ./companion/indexer_daemon/
-COPY companion/rag/ ./companion/rag/
+COPY src/companion/ ./companion/
+COPY src/companion/indexer_daemon/ ./companion/indexer_daemon/
+COPY src/companion/rag/ ./companion/rag/
 
 # Create directories for data
 RUN mkdir -p /data/vectors

README.md

@@ -39,6 +39,7 @@ A fully local, privacy-first AI companion trained on your Obsidian vault. Combin
 - Node.js 18+ (for UI)
 - Ollama running locally
 - RTX 5070 or equivalent (12GB+ VRAM for fine-tuning)
+- See [GPU Compatibility Guide](docs/gpu-compatibility.md) for RTX 50-series setup
 
 ### Installation
 
@@ -76,7 +77,7 @@ See [docs/config.md](docs/config.md) for full configuration reference.
 
 **Terminal 1 - Backend:**
 ```bash
-python -m uvicorn companion.api:app --host 0.0.0.0 --port 7373
+python -m companion.api
 ```
 
 **Terminal 2 - Frontend:**
@@ -139,8 +140,9 @@ python -m companion.forge.cli reload ~/.companion/training/final
 | `companion.config` | Configuration management | [docs/config.md](docs/config.md) |
 | `companion.rag` | RAG engine (chunk, embed, search) | [docs/rag.md](docs/rag.md) |
 | `companion.forge` | Fine-tuning pipeline | [docs/forge.md](docs/forge.md) |
-| `companion.api` | FastAPI backend | [docs/api.md](docs/api.md) |
+| `companion.api` | FastAPI backend | This README |
 | `ui/` | React frontend | [docs/ui.md](docs/ui.md) |
+| **GPU Setup** | RTX 50-series compatibility | [docs/gpu-compatibility.md](docs/gpu-compatibility.md) |
 
 ## Project Structure
 

config.json

@@ -55,7 +55,7 @@
   "rag": {
     "embedding": {
       "provider": "ollama",
-      "model": "mxbai-embed-large",
+      "model": "mxbai-embed-large:335m",
       "base_url": "http://localhost:11434",
       "dimensions": 1024,
       "batch_size": 32

docker-compose.yml

@@ -1,4 +1,4 @@
-version: '3.8'
+version: "3.8"
 
 services:
   companion-api:
@@ -20,7 +20,13 @@ services:
       - companion-network
     restart: unless-stopped
     healthcheck:
-      test: ["CMD", "python", "-c", "import requests; requests.get('http://localhost:7373/health')"]
+      test:
+        [
+          "CMD",
+          "python",
+          "-c",
+          "import requests; requests.get('http://localhost:7373/health')",
+        ]
       interval: 30s
       timeout: 10s
       retries: 3
@@ -34,7 +40,7 @@ services:
     volumes:
       - ./config.json:/app/config.json:ro
       - companion-data:/data
-      - /home/san/KnowledgeVault:/vault:ro  # Mount Obsidian vault as read-only
+      - ./sample-data/Default:/app/sample-data/Default:ro # Mount Obsidian vault as read-only
     environment:
       - COMPANION_CONFIG=/app/config.json
       - COMPANION_DATA_DIR=/data

docs/forge.md

@@ -193,22 +193,24 @@ new_path = reload_model(
 
 ```bash
 # Extract training data
-companion.forge.cli extract [--output PATH]
+python -m companion.forge.cli extract [--output PATH]
 
 # Train model
-companion.forge.cli train \
-  [--data PATH] \
-  [--output PATH] \
-  [--epochs N] \
-  [--lr FLOAT]
+python -m companion.forge.train \
+  --data PATH \
+  --output-dir PATH \
+  --epochs N \
+  --lr FLOAT
 
 # Check model status
-companion.forge.cli status
+python -m companion.forge.cli status
 
 # Reload model
-companion.forge.cli reload MODEL_PATH [--no-backup]
+python -m companion.forge.cli reload MODEL_PATH [--no-backup]
 ```
 
+**Note:** Use `--output-dir` (or `--output`) to specify the training output directory.
+
 ## Training Tips
 
 **Dataset Size:**
@@ -243,6 +245,11 @@ Use `gradient_accumulation_steps` to increase effective batch size.
 
 ## Troubleshooting
 
+**GPU Not Detected / CUDA Not Available**
+- See [GPU Compatibility Guide](gpu-compatibility.md)
+- Common issue on RTX 50-series: Install CUDA-enabled PyTorch: `pip install torch==2.5.1+cu121 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121`
+- Verify: `python -c "import torch; print(torch.cuda.is_available())"`
+
 **CUDA Out of Memory**
 - Reduce `lora_rank` to 8
 - Reduce `batch_size` to 2

docs/gpu-compatibility.md

@@ -0,0 +1,97 @@
+# GPU Compatibility Guide
+
+## RTX 50-Series (Blackwell) Compatibility Notice
+
+### Issue
+NVIDIA RTX 50-series GPUs (RTX 5070, 5080, 5090) use CUDA capability `sm_120` (Blackwell architecture). PyTorch stable releases (up to 2.5.1) only officially support up to `sm_90` (Hopper/Ada).
+
+**Warning you'll see:**
+```
+NVIDIA GeForce RTX 5070 with CUDA capability sm_120 is not compatible with the current PyTorch installation.
+The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
+```
+
+### Current Status
+- ✅ PyTorch detects the GPU
+- ✅ CUDA operations generally work
+- ⚠️ Some operations may fail or fall back to CPU
+- ⚠️ Performance may not be optimal
+
+### Workarounds
+
+#### Option 1: Use PyTorch Nightly (Recommended for RTX 50-series)
+```bash
+pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124
+```
+
+#### Option 2: Use Current Stable with Known Limitations
+Many workloads work fine despite the warning. Test your specific use case.
+
+#### Option 3: Wait for PyTorch 2.7
+Full sm_120 support is expected in the next stable release.
+
+### Installation Steps for KV-RAG with GPU
+
+1. **Install CUDA-enabled PyTorch:**
+   ```bash
+   pip install torch==2.5.1+cu121 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
+   ```
+
+2. **Install unsloth without dependencies:**
+   ```bash
+   pip install unsloth --no-deps
+   pip install unsloth_zoo
+   ```
+
+3. **Install remaining training dependencies:**
+   ```bash
+   pip install bitsandbytes accelerate peft transformers datasets trl
+   ```
+   Note: Skip `xformers` as it may overwrite torch. Unsloth works without it.
+
+### Verify GPU is Working
+
+```python
+import torch
+print(f"CUDA available: {torch.cuda.is_available()}")
+print(f"GPU: {torch.cuda.get_device_name(0)}")
+print(f"CUDA version: {torch.version.cuda}")
+```
+
+### Ollama GPU Status
+
+Ollama runs **natively on Windows** and uses GPU automatically when available:
+- Check with: `nvidia-smi` (look for `ollama.exe` processes)
+- Embedding model (`mxbai-embed-large:335m`) runs on GPU
+- Chat models also use GPU when loaded
+
+### Forge Training GPU Status
+
+The training script uses `unsloth` + `trl` for QLoRA fine-tuning:
+- Requires CUDA-enabled PyTorch
+- Optimized for 12GB VRAM (RTX 5070)
+- Uses 4-bit quantization + LoRA adapters
+- See `src/companion/forge/train.py` for implementation
+
+### Troubleshooting
+
+**Issue:** `CUDA available: False` after installation
+**Fix:** PyTorch was overwritten by a package dependency. Reinstall:
+```bash
+pip install torch==2.5.1+cu121 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 --force-reinstall
+```
+
+**Issue:** `xformers` overwrites torch
+**Fix:** Skip xformers or install matching wheel:
+```bash
+# Skip for now - unsloth works without it
+# Or install specific version matching your torch
+pip install xformers==0.0.28.post3 --index-url https://download.pytorch.org/whl/cu121
+```
+
+### References
+
+- [PyTorch CUDA Compatibility](https://pytorch.org/get-started/locally/)
+- [NVIDIA CUDA Capability Matrix](https://developer.nvidia.com/cuda-gpus)
+- [Unsloth Documentation](https://github.com/unsloth/unsloth)
+- [RTX 50-Series Architecture](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/)

pyproject.toml

@@ -37,7 +37,7 @@ train = [
     "trl>=0.7.0",
 ]
 
-[tool.hatchling]
+[tool.hatch.build.targets.wheel]
 packages = ["src/companion"]
 
 [build-system]

src/companion/api.py

@@ -8,7 +8,7 @@ from contextlib import asynccontextmanager
 from typing import AsyncGenerator
 
 import httpx
-from fastapi import FastAPI, HTTPException
+from fastapi import FastAPI, HTTPException, APIRouter
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 from sse_starlette.sse import EventSourceResponse
@@ -37,6 +37,12 @@ class ChatResponse(BaseModel):
     sources: list[dict] | None = None
 
 
+class ReloadModelRequest(BaseModel):
+    """Model reload request."""
+
+    model_path: str
+
+
 # Global instances
 config: Config
 vector_store: VectorStore
@@ -70,8 +76,7 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
     orchestrator = ChatOrchestrator(
         config=config,
         search_engine=search_engine,
-        memory=memory,
-        http_client=http_client,
+        session_memory=memory,
     )
 
     yield
@@ -99,8 +104,11 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
+# Create API router with /api prefix
+api_router = APIRouter(prefix="/api")
 
-@app.get("/health")
+
+@api_router.get("/health")
 async def health_check() -> dict:
     """Health check endpoint."""
     return {
@@ -110,7 +118,7 @@ async def health_check() -> dict:
     }
 
 
-@app.post("/chat")
+@api_router.post("/chat")
 async def chat(request: ChatRequest) -> EventSourceResponse:
     """Chat endpoint with SSE streaming."""
     if not request.message.strip():
@@ -168,7 +176,7 @@ async def chat(request: ChatRequest) -> EventSourceResponse:
     )
 
 
-@app.get("/sessions/{session_id}/history")
+@api_router.get("/sessions/{session_id}/history")
 async def get_session_history(session_id: str) -> dict:
     """Get conversation history for a session."""
     history = memory.get_history(session_id)
@@ -185,13 +193,7 @@ async def get_session_history(session_id: str) -> dict:
     }
 
 
-class ReloadModelRequest(BaseModel):
-    """Model reload request."""
-
-    model_path: str
-
-
-@app.post("/admin/reload-model")
+@api_router.post("/admin/reload-model")
 async def reload_model_endpoint(request: ReloadModelRequest) -> dict:
     """Reload the model with a new fine-tuned version (admin only)."""
     from pathlib import Path
@@ -214,6 +216,10 @@ async def reload_model_endpoint(request: ReloadModelRequest) -> dict:
         raise HTTPException(status_code=500, detail=f"Failed to reload model: {e}")
 
 
+# Include the API router
+app.include_router(api_router)
+
+
 if __name__ == "__main__":
     import uvicorn
 

src/companion/forge/train.py

@@ -211,7 +211,9 @@ def main():
         "--data", type=Path, required=True, help="Path to training data JSONL"
     )
     parser.add_argument(
+        "--output-dir",
         "--output",
+        dest="output",
         type=Path,
         default=Path("~/.companion/training"),
         help="Output directory",

src/companion/rag/chunker.py

@@ -28,16 +28,27 @@ class ChunkingRule:
 
 
 def sliding_window_chunks(text: str, chunk_size: int, chunk_overlap: int) -> List[str]:
-    words = text.split()
-    if not words:
+    """Split text into chunks based on character count (not word count)."""
+    if not text:
         return []
 
     chunks = []
     step = chunk_size - chunk_overlap
-    for i in range(0, len(words), step):
-        chunk_words = words[i : i + chunk_size]
-        chunks.append(" ".join(chunk_words))
-        if i + chunk_size >= len(words):
+    start = 0
+    text_len = len(text)
+
+    while start < text_len:
+        end = min(start + chunk_size, text_len)
+        # Try to break at word boundary
+        if end < text_len:
+            # Look for whitespace to break at
+            while end > start and not text[end].isspace():
+                end -= 1
+            if end == start:  # No good break found, force cut
+                end = min(start + chunk_size, text_len)
+        chunks.append(text[start:end].strip())
+        start += step
+        if end >= text_len:
             break
 
     return chunks