M.I.M.I.R - Multi-agent Intelligent Memory & Insight Repository

# NornicDB Docker Build Fix ## Issues Fixed ### 1. ✅ Embedder Model Loading Failure (CRITICAL) **Error**: `llama_model_load: error loading model: vector::_M_range_check: __n (which is 1) >= this->size() (which is 1)` **Root Cause**: llama.cpp version b4785 has a bug that prevents loading certain GGUF models like bge-m3.gguf **Solution**: Upgraded llama.cpp from b4785 → b7285 (latest stable, Dec 2025) **Files Changed**: - `nornicdb/docker/Dockerfile.llama-cuda`: Updated `ARG LLAMA_VERSION=b7285` - `nornicdb/docker/Dockerfile.amd64-cuda`: Updated base image reference to `b7285` ### 2. ✅ Build Version Verification **Problem**: No way to verify which Docker image version is actually running **Solution**: Added git commit hash and build timestamp to startup logs **Format**: `🚀 Starting NornicDB v0.1.0-abc1234 (built: 20250106-123456)` **Files Changed**: - `nornicdb/cmd/nornicdb/main.go`: Enhanced version display logic - `nornicdb/docker/Dockerfile.amd64-cuda`: Added commit hash to build ldflags ## Rebuild Instructions ### Step 1: Rebuild llama-cuda-libs base image (one-time, ~15 min) ```powershell cd C:\Users\timot\Documents\GitHub\Mimir\nornicdb # Build the new llama.cpp b7285 static library docker build -f docker/Dockerfile.llama-cuda -t timothyswt/llama-cuda-libs:b7285 . # (Optional) Push to registry for reuse docker push timothyswt/llama-cuda-libs:b7285 ``` ### Step 2: Rebuild NornicDB image (~2 min with cached base) ```powershell # Full build with BGE model embedded and Heimdall docker build -f docker/Dockerfile.amd64-cuda ` --build-arg EMBED_MODEL=true ` -t timothyswt/nornicdb-amd64-cuda-bge-heimdall:v1.0.1 . # Or build without embedding model (BYOM) docker build -f docker/Dockerfile.amd64-cuda ` -t timothyswt/nornicdb-amd64-cuda:v1.0.1 . ``` ### Step 3: Test the fix ```powershell # Stop old container docker stop nornicdb docker rm nornicdb # Start new container docker run -d --name nornicdb ` --gpus all ` -p 7474:7474 -p 7687:7687 ` -v C:\Users\timot\Documents\GitHub\Mimir\data\nornicdb:/data ` -v C:\Users\timot\Documents\GitHub\Mimir\models:/app/models ` timothyswt/nornicdb-amd64-cuda-bge-heimdall:v1.0.1 # Check logs - should see successful model loading docker logs nornicdb -f ``` ## Expected Log Output (Success) ``` 🚀 Starting NornicDB v0.1.0-a1b2c3d (built: 20250106-143022) Data directory: /data Bolt protocol: bolt://localhost:7687 HTTP API: http://localhost:7474 🧠 Loading local embedding model: /app/models/bge-m3.gguf GPU layers: -1 (-1 = auto/all) ✅ Model loaded: bge-m3 (1024 dimensions) 🔥 Model warmup enabled: every 5m0s ✓ Embeddings enabled: local GGUF (bge-m3, 1024 dims) ``` ## Verification 1. **Check version includes commit hash**: ``` 🚀 Starting NornicDB v0.1.0-a1b2c3d (built: 20250106-143022) ``` 2. **Confirm model loads without error**: ``` ✅ Model loaded: bge-m3 (1024 dimensions) ``` 3. **Test embeddings**: - Open http://localhost:7474 - Click "Regenerate Embeddings" - Should return success (not 503) ## Rollback (if needed) ```powershell # Revert to old working image (Mac build) docker stop nornicdb docker rm nornicdb docker run -d --name nornicdb --gpus all -p 7474:7474 -p 7687:7687 ` -v C:\Users\timot\Documents\GitHub\Mimir\data\nornicdb:/data ` timothyswt/nornicdb-amd64-cuda-bge-heimdall:v1.0.0 ``` ## Technical Details ### Why this fixes the embedder issue: 1. **Old llama.cpp b4785** (several months old): - Had a bug in GGUF model loading that caused vector bounds check failures - Specific to certain model architectures like BGE-M3 - Manifested as: `vector::_M_range_check: __n (which is 1) >= this->size() (which is 1)` 2. **New llama.cpp b7285** (Dec 2025): - Fixed GGUF loading bugs - Improved model compatibility - Better error messages - CUDA optimizations ### Why build timestamp matters: - Docker image tags (`:latest`, `:v1.0.0`) can be ambiguous - Commit hash proves exactly which code is running - Build timestamp shows when image was created - Enables debugging: "Was this built before or after the fix?" ## Related Files - `nornicdb/docker/Dockerfile.llama-cuda` - llama.cpp static library builder - `nornicdb/docker/Dockerfile.amd64-cuda` - Main NornicDB CUDA build - `nornicdb/cmd/nornicdb/main.go` - Startup logging with version info - `nornicdb/pkg/embed/local_gguf.go` - Local GGUF embedder (calls llama.cpp) - `nornicdb/pkg/localllm/llama_windows.go` - llama.cpp CGO bindings ## Notes - The old `timothyswt/llama-cuda-libs:b4785` image can be kept for now (doesn't hurt) - New builds will automatically use b7285 due to Dockerfile changes - Build time ~17 minutes total: - llama-cuda-libs: ~15 min (one-time) - nornicdb: ~2 min (uses cached llama-cuda-libs)