Local C++ RAG server for indexing, enriching, and searching codebases. All models run locally via llama.cpp — no cloud required.
Wick is a derivative of Wax — ported from Swift to C++20, extended with an HTTP server, LLM enrichment pipeline, structured knowledge store, and MCP bridge for IDE integration. See NOTICE for full attribution.
Work in progress — APIs, storage format, and configuration may change.
git submodule update --init --recursive
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
ctest --test-dir build --output-on-failure -C ReleaseWick uses llama.cpp for fact enrichment and answer generation. Start a llama-server with any GGUF model:
llama-server --model /path/to/model.gguf --host 127.0.0.1 --port 8004 --ctx-size 8192 --n-gpu-layers 99export WAXCPP_LLAMA_CPP_ROOT=/path/to/llama-cpp
export WAXCPP_GENERATION_MODEL=/path/to/model.gguf
./build/bin/waxcpp_rag_serverThe server starts on http://127.0.0.1:8080 and communicates via JSON-RPC 2.0.
curl -s -X POST http://127.0.0.1:8080/ \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"index.start","params":{
"repo_root":"/path/to/your/project",
"include_extensions":[".h",".cpp",".hpp"],
"enrich_regex":true,
"enrich_llm":true
}}'# Full-text search
curl -s -X POST http://127.0.0.1:8080/ \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"recall","params":{"query":"how does hashing work","top_k":5}}'
# RAG-augmented answer
curl -s -X POST http://127.0.0.1:8080/ \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"answer.generate","params":{"query":"Explain the actor lifecycle"}}' IDE / Agent ───> MCP bridge (Node.js, stdio)
│
JSON-RPC / HTTP :8080
│
waxcpp_rag_server
┌────┴────┐
BM25/FTS5 Structured facts (EAV)
(SQLite) (in-memory, persisted)
└────┬────┘
.mv2s store
(append-only binary + WAL)
| Method | Description |
|---|---|
remember |
Store a text chunk with metadata |
recall |
Hybrid search (BM25 + optional vector) |
answer.generate |
RAG-augmented answer via local LLM |
index.start |
Start background indexing of a directory |
index.status |
Poll indexing progress (state, phase, metrics) |
index.stop |
Cancel a running index job |
fact.search |
Search structured entity-attribute-value facts |
flush |
Force-flush pending writes to disk |
blueprint.read |
Read an exported UE5 Blueprint JSON |
blueprint.write |
Write modified Blueprint JSON (with backup) |
blueprint.import |
Trigger Blueprint reimport into .uasset |
| Parameter | Type | Default | Description |
|---|---|---|---|
repo_root |
string | required | Path to the directory to index |
resume |
bool | false |
Skip unchanged files using checkpoint manifests |
include_extensions |
string[] | [".h",".cpp",".hpp",".inl",".inc"] |
File extensions to scan |
exclude_dirs |
string[] | [] |
Directory names to skip |
enrich_regex |
bool | false |
Extract facts via regex (class hierarchies, macros) |
enrich_llm |
bool | false |
Extract facts via local LLM |
target_tokens |
int | 400 |
Target chunk size in tokens |
flush_every_chunks |
int | 128 |
Commit/checkpoint cadence |
ingest_batch_size |
int | 1 |
Chunks buffered before batch apply |
max_files |
int | 0 |
Cap on scanned files (0 = no cap) |
max_chunks |
int | 0 |
Cap on ingested chunks (0 = no cap) |
max_ram_mb |
int | 0 |
Soft RSS cap in MB (0 = disabled) |
The mcp/ directory contains a Node.js MCP server that connects editors (VS Code, Claude Code) to the Wick server via stdio.
Available tools: wax_recall, wax_remember, wax_fact_search, wax_blueprint_read, wax_blueprint_write.
cd mcp && npm installMCP config example (mcp.json):
{
"servers": {
"wax": {
"type": "stdio",
"command": "node",
"args": ["/path/to/wick/mcp/wax_mcp_server.js"],
"env": { "WAX_URL": "http://127.0.0.1:8080" }
}
}
}Wick has built-in support for Unreal Engine 5 projects:
| Source type | Extensions | Enrichment |
|---|---|---|
| C++ source | .h, .hpp, .cpp, .inl, .inc |
Regex (UCLASS/UPROPERTY/UFUNCTION macros, class hierarchies) + optional LLM |
| Blueprints | .bpl_json (exported via BlueprintGraphExport commandlet) |
LLM (entity-attribute-value facts: calls, variables, events, inheritance) |
Example scripts for indexing UE5 projects are in ue5_sandbox/.
| Variable | Description |
|---|---|
WAXCPP_LLAMA_CPP_ROOT |
Path to llama.cpp binaries (required) |
WAXCPP_GENERATION_MODEL |
Path to generation GGUF model |
WAXCPP_SERVER_CONFIG |
Path to server-runtime.json |
WAXCPP_SERVER_LOG |
Enable server logging (1) |
WAXCPP_RECALL_LOG |
File path for recall request/response log |
WAXCPP_AUTO_FLUSH_INTERVAL_MS |
Auto-flush interval in ms (default 30000) |
| Variable | Description |
|---|---|
WAXCPP_LLAMA_GEN_ENDPOINT |
Generation endpoint (default http://127.0.0.1:8004/completion) |
WAXCPP_LLAMA_API_KEY |
API key for llama-server |
WAXCPP_LLAMA_GEN_TIMEOUT_MS |
Request timeout in ms |
WAXCPP_LLAMA_GEN_MAX_RETRIES |
Retry count on failures |
| Variable | Description |
|---|---|
WAXCPP_ENRICH_LLM_LOG |
Verbose enrichment logging (1) |
WAXCPP_ENRICH_LLM_MAX_TOKENS |
Max tokens for enrichment responses (default 1024) |
| Variable | Description |
|---|---|
WAXCPP_LLAMA_EMBED_ENDPOINT |
Embedding endpoint URL |
WAXCPP_LLAMA_EMBED_DIMS |
Expected embedding dimension |
WAXCPP_LLAMA_EMBED_TIMEOUT_MS |
Request timeout in ms |
WAXCPP_LLAMA_EMBED_MAX_RETRIES |
Retry count |
WAXCPP_LLAMA_EMBED_MAX_BATCH_CONCURRENCY |
Max parallel workers |
| Variable | Description |
|---|---|
WAXCPP_RAG_MAX_SNIPPETS |
Max snippets in RAG context |
WAXCPP_RAG_SEARCH_TOP_K |
Top-K results for search |
WAXCPP_ORCH_INGEST_CONCURRENCY |
Parallel ingest workers |
WAXCPP_ORCH_INGEST_BATCH_SIZE |
Ingest batch size |
.
├── include/waxcpp/ # Public headers (core library)
├── src/ # Core library implementation
│ ├── core/ # Storage, WAL, SHA256
│ ├── text/ # FTS5, structured memory
│ ├── vector/ # USearch vector engine
│ ├── rag/ # Search, embeddings, scoring, chunking
│ └── orchestrator/ # Memory orchestrator
├── server/ # HTTP RAG server (Poco-based)
├── mcp/ # MCP bridge for IDE integration
├── tests/ # Unit, parity, and integration tests
├── fixtures/ # Cross-language parity fixtures
├── tools/ # mv2s_analyzer, mv2s_vacuum
├── scripts/ # Submodule verification, CI helpers
├── ue5_sandbox/ # UE5-specific indexing/export scripts
├── third_party/ # Git submodules (usearch, sqlite, googletest, poco)
├── manifest/ # Artifact manifests
└── docs/ # Additional documentation
All third-party dependencies are managed via git submodules under third_party/.
See: .gitmodules, submodules.lock, scripts/verify_submodules.py.
# Run policy test
ctest --test-dir build --output-on-failure -R waxcpp_verify_submodules_policy_test
# Strict CI checks
python3 scripts/verify_submodules.py --require-gitlinks-present
python3 scripts/verify_submodules.py --enforce-pin-required