Installing _AugmentedIntelligence on Windows
This is the primary, fully-featured build: the Visual Studio C++ project with the Dear ImGui + DirectX 11 dashboard, the typed-command REPL, the perception bus, RSI agents, Lua extensibility, media/legal databases, and optional NVIDIA GPU acceleration.
There are two paths:
- Guided installer —
installer\install.ps1checks prerequisites, installs the lightweight ones, builds vcpkg ports + llama.cpp, compiles the solution, and deploys a runnable tree. Recommended. - Manual build — open the solution in Visual Studio and build a configuration yourself. Use this when you want fine control (e.g. a specific CPU/GPU variant).
1. System requirements
| Minimum | Recommended | |
|---|---|---|
| OS | Windows 10 21H2 x64 | Windows 11 x64 |
| RAM | 16 GB | 32 GB+ (linking the full binary is memory-heavy) |
| Disk | ~30 GB free | 60 GB+ (deps, llama builds, models, databases) |
| CPU | x64 with SSE2 | x64 with AVX2 or AVX-512 |
| GPU | none (CPU build) | NVIDIA (GTX 10-series+), 8 GB+ VRAM for local LLMs |
GPU is optional — see Build variants. A no-GPU, no-AVX build runs on any x64 machine.
2. Prerequisites
Large items (top of the table) are installed manually; the installer's prerequisites phase handles the small ones (Git, Python, OpenSSH, VC++ redist).
| Component | Version | Notes / link |
|---|---|---|
| Visual Studio | 2022 17.14+ or 2026 | Workload: Desktop development with C++, MSVC toolset v145. https://visualstudio.microsoft.com/downloads/ |
| NVIDIA CUDA Toolkit | 12.5 | GPU builds only. Integrate with VS during install. https://developer.nvidia.com/cuda-downloads |
| MySQL Server | 8.0 | For music/video/legal libraries. https://dev.mysql.com/downloads/ |
| MySQL Connector/C++ | 8.0 + 9.0 | Headers (8.0) + libs (9.0). Both paths are referenced by the project. |
| vcpkg | latest | Third-party C/C++ ports (default root E:\vcpkg\vcpkg). https://github.com/microsoft/vcpkg |
| Python | 3.12 | Embedded Python + import/utility scripts. |
| Git | latest | Cloning sources. |
| Ollama | latest | Optional external LLM backend for chat/RSI. https://ollama.com |
vcpkg ports
The installer runs these for you (-Phase vcpkg). Triplet: x64-windows.
curl[ssl,schannel,sspi] ffmpeg[avcodec,avdevice,avfilter,avformat,swresample,swscale]
giflib kissfft leptonica libarchive[bzip2,crypto,libxml2,lz4,lzma] libjpeg-turbo
libpng libxml2[iconv,lzma,zlib] lua lz4 openssl rtaudio tesseract
tiff[jpeg,lzma,zip] zlib boost-python libsndfile libwebp openjpeg bzip2 libiconv
Prebuilt libraries under the source root
A few heavy libraries are not from vcpkg and must be present under C:\_AugmentedIntelligence\_src (the AiSrcRoot):
opencv_world490.lib/opencv_world490.dll(OpenCV 4.9.0)tensorflow.lib/tensorflow.dll(TensorFlow C API)whisper.lib/whisper.dll(whisper.cpp)quirc.lib,sndfile.libthird_party\litert\— LiteRT / TFLite C (tensorflowlite_c.dll) ships in the repo.
3. Quick start (guided installer)
From a Developer PowerShell (or normal PowerShell with VS on PATH), in the project root:
# 1. See what's present / missing — changes nothing
.\installer\install.ps1 -CheckOnly
# 2. Full install: prerequisites, vcpkg, build, deploy, post-install
.\installer\install.ps1 -Phase all
# 2b. Include the CUDA llama.cpp build and copy model trees
.\installer\install.ps1 -Phase all -IncludeLlamaBuild -IncludeModels
Run specific phases only:
.\installer\install.ps1 -Phase build,deploy -SkipPrerequisites
Phases: check → prerequisites → vcpkg → llama (opt-in) → build → deploy → postinstall.
Key parameters / environment overrides:
| Parameter | Env var | Default |
|---|---|---|
-RuntimeRoot |
AI_RUNTIME_ROOT |
C:\_AugmentedIntelligence |
-AiSrcRoot |
AI_SRC_ROOT |
C:\_AugmentedIntelligence\_src |
-VcpkgRoot |
AI_VCPKG_ROOT |
E:\vcpkg\vcpkg |
-CudaVersion |
AI_CUDA_VERSION |
12.5 |
-Configuration |
— | Release |
-IncludeLlamaBuild |
— | off (adds the llama phase) |
-IncludeModels |
— | off (copies tensorflow_models, whisper_models, ggml_models) |
-NonInteractive |
— | off (no prompts; for CI) |
The deploy phase copies the exe, lua_scripts\, scripts\, third_party\litert\, and all transitive runtime DLLs (resolved with dumpbin /dependents) into RuntimeRoot.
4. Manual build (Visual Studio)
- Place sources. Clone/copy the project so the tree looks like:
C:\_AugmentedIntelligence\_src\
_AugmentedIntelligence_v11.0 - Typed Commands\ <- this project (.sln)
llama.cpp\ <- llama.cpp source
opencv_world490.lib, tensorflow.lib, whisper.lib, ...
-
Install vcpkg ports (see list above) and run
vcpkg integrate install. -
Set up MySQL — install Server 8.0, create a user, and create the databases used by the library importers (
music_library,video_library, and the legalus_codeschema). Put credentials insettings.txt(see Configuration). -
Build llama.cpp — see Build variants for the exact CMake flags per variant.
-
Open
_AugmentedIntelligence.slnin Visual Studio, pick a configuration (e.g.Release-GPU-AVX2 | x64), and Build. Or from a Developer prompt:
powershell
msbuild "_AugmentedIntelligence.sln" /p:Configuration=Release /p:Platform=x64 /m
Output: x64\<Configuration>\_AugmentedIntelligence.exe.
5. MySQL databases (optional features)
The core runtime works without MySQL; the media and legal features need it.
- music_library — one table per artist (populated from FLAC via the Python importer).
- video_library —
4k_blu_rays,blu_rays,dvds,television(LLM-enriched movie/TV metadata). - us_code — legal corpus (58-jurisdiction registry).
Import scripts live under the project (music_library\, video_library\). They read credentials from settings.txt, so you never pass the password on the command line.
6. Build variants
The solution ships these x64 configurations (GPU is opt-in; SIMD level is baked into both the app and its matching llama.cpp build):
| Configuration | GPU (CUDA) | App SIMD | llama.cpp build dir |
|---|---|---|---|
Release / Debug |
✅ | (default) | llama_build\ |
Release-GPU-AVX512 |
✅ | AVX-512 | llama_build\ |
Release-GPU-AVX2 |
✅ | AVX2 | llama_build_cuda_avx2\ |
Release-CPU-AVX2 |
❌ | AVX2 | llama_build_cpu_avx2\ |
Release-CPU-NoAVX |
❌ | SSE2 (no AVX) | llama_build_cpu_noavx\ |
CUDA is controlled by the UseCuda MSBuild property (true for the GPU configs). A CPU config links no CUDA libraries, so the resulting exe has no nvcuda.dll dependency and runs on machines with no NVIDIA driver.
Build the matching llama.cpp for a variant (all use -G "Visual Studio 17 2022" -A x64 -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_SERVER=OFF -DLLAMA_BUILD_TESTS=OFF):
# GPU + AVX-512 (the installer's -IncludeLlamaBuild builds this one)
cmake -S ...\llama.cpp -B ...\llama_build -DGGML_CUDA=ON -DGGML_NATIVE=ON
# GPU + AVX2
cmake -S ...\llama.cpp -B ...\llama_build_cuda_avx2 -DGGML_CUDA=ON -DGGML_NATIVE=OFF -DGGML_AVX2=ON -DGGML_AVX=ON -DGGML_FMA=ON -DGGML_AVX512=OFF
# CPU + AVX2
cmake -S ...\llama.cpp -B ...\llama_build_cpu_avx2 -DGGML_CUDA=OFF -DGGML_NATIVE=OFF -DGGML_AVX2=ON -DGGML_AVX=ON -DGGML_FMA=ON -DGGML_AVX512=OFF
# CPU + no AVX (SSE2 baseline; runs anywhere)
cmake -S ...\llama.cpp -B ...\llama_build_cpu_noavx -DGGML_CUDA=OFF -DGGML_NATIVE=OFF -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DGGML_AVX512=OFF
# then, for any of the above:
cmake --build <build-dir> --config Release --target llama llama-common llama-common-base ggml ggml-base ggml-cpu
⚠️
GGML_NATIVE=ONcompiles for the build machine's CPU (may emit AVX-512). For a binary you'll run on older/other machines, use the explicit AVX2 or no-AVX variant above.
Full rationale in llama_cpp_cpu_ram_cuda.md.
7. Configuration
MySQL is the primary configuration store (ai_settings). Put MySQL connection credentials in settings.txt in the runtime root (bootstrap only). All other settings (feature flags, LLM hostnames, matrix mode auto/avx512/avx/scalar, etc.) load from and save to MySQL after first connect.
On first run with an empty ai_settings store, a full legacy settings.txt is migrated into MySQL automatically. Details: MySQLConfiguration.md.
Keep bootstrap settings.txt out of source control; it holds secrets. Optional: settings_write_file_mirror=true or CLI --settings-file-mirror for a full local cache.
8. LLM backend
LLM-driven features (character chat, RSI agents, Lua orchestrators, video metadata enrichment) need a language-model backend:
- In-process llama.cpp — linked from the
llama_build*directory of your chosen variant. - Ollama (recommended for GPU offload) — install Ollama, then pull models, e.g.:
powershell
ollama pull gemma3:12b
ollama pull llama3:8b-instruct-q8_0
The app talks to Ollama at http://localhost:11434. Set the model name in settings.txt.
9. Run
# From the deployed runtime root:
C:\_AugmentedIntelligence\_AugmentedIntelligence.exe
# Or straight from the build output:
.\x64\Release-GPU-AVX2\_AugmentedIntelligence.exe
At the typed-command prompt, useful first commands:
hardware # show detected CPU/GPU + matrix-acceleration status
help # command overview
web server start # browser dashboard
10. Remote / mobile pairing
The postinstall phase opens firewall rules for TCP 7710 and 7720. To pair the Android companion:
remote commands start 7710 YOUR_TOKEN
android audio start 7711 YOUR_TOKEN
web server start
See Pixel4-Install.md.
11. Packaging an installer
To build a redistributable payload/installer (bundles the exe + all runtime DLLs from installer\payload\Dependencies\):
.\installer\prepare-payload.ps1 # stage exe + dependencies
.\installer\build-installer.ps1 # produce the installer package
12. Troubleshooting
nvcuda.dll is missing on launch.
The machine has no NVIDIA driver but you ran a GPU build. Use a CPU configuration (Release-CPU-AVX2 or Release-CPU-NoAVX) — those link no CUDA. This is a load-time failure a runtime check can't fix; the fix is the build variant.
Illegal instruction / crash at startup on an older PC.
The binary uses a SIMD level the CPU lacks. Use Release-CPU-NoAVX (SSE2) or Release-CPU-AVX2, and the matching no-AVX/AVX2 llama build.
MSBuild: unresolved ggml_backend_cuda_* symbols in a CPU build.
The CPU config is pointing at a CUDA-enabled llama build. Build the matching CPU-only llama_build_cpu_* directory (§6) so LlamaBuildDir resolves to CUDA-free libs.
MSBuild not found from the installer.
Install Visual Studio with the C++ desktop workload; the installer locates MSBuild via vswhere.
Missing opencv_world490.lib / tensorflow.lib / llama.lib (installer [--]).
These are the large prebuilt artifacts under AiSrcRoot. Populate them and, for llama, run -Phase llama -IncludeLlamaBuild.
MySQL connection errors.
Verify Server 8.0 is running and the credentials/database names in settings.txt are correct. The library features are optional — the core runtime starts without them.
Related documents
- Ubuntu-Install.md — Linux (console/web) port
- Pixel4-Install.md — Android companion
- llama_cpp_cpu_ram_cuda.md — llama.cpp memory / CUDA tuning
installer\install.ps1— guided Windows installer (source of truth for phases & defaults)