# RAG (Retrieval-Augmented Generation) [Top](../Readme.md) | [Architecture](guide_architecture.md) | [MMA](guide_mma.md) | [Tools & IPC](guide_tools.md) | [Simulations](guide_simulations.md) --- ## Overview Manual Slop integrates Retrieval-Augmented Generation (RAG) to extend the AI's working context beyond the explicit file list. When a project is RAG-enabled, the system maintains a vector index of file content; AI calls can retrieve semantically similar fragments at query time and prepend them to the prompt. The RAG implementation is pluggable: the vector store, the embedding provider, and the chunking strategy are all configurable per project. The default backend is **ChromaDB** (local persistent), the default embedding is **Gemini Embedding 001** (cloud), and the default chunking is **character-based with overlap** (with **AST-aware chunking** for Python files when enabled). This guide covers: 1. **Architecture** — Where RAG fits in the dispatch pipeline 2. **Components** — `RAGEngine`, embedding providers, vector store 3. **Data Flow** — Indexing, query, retrieval, injection 4. **Configuration** — `RAGConfig` schema and TOML settings 5. **Verification** — Test infrastructure and known edge cases --- ## Architecture RAG sits between the project's tracked files and the AI provider's input prompt. It is **not** an internal AI call — it is a pre-processing step that augments `md_content` before the provider sees it. ``` ┌─────────────────────────────────┐ │ AppController / ConductorEngine │ │ (caller of ai_client.send) │ └────────────┬────────────────────┘ │ constructs RAGEngine once per project ▼ ┌────────────────────────────────────────────┐ │ RAGEngine │ │ ├─ EmbeddingProvider (Local or Gemini) │ │ ├─ VectorStore (ChromaDB persistent) │ │ └─ Chunkers (_chunk_text, _chunk_code) │ └────────────┬───────────────────────────────┘ │ on every ai_client.send() call: │ rag_engine.search(user_message) -> fragments ▼ ┌────────────────────────────────────────────┐ │ ai_client.send(rag_engine=...) │ │ injects [RETRIEVED CONTEXT] block │ │ into md_content before provider call │ └────────────────────────────────────────────┘ ``` **Lifecycle**: - The `AppController` constructs a single `RAGEngine` per project load (lazily, when the project is first opened or when a RAG-related setting changes). - The `RAGEngine` is passed through to `ai_client.send()` for every AI call from the main discussion flow. - For Tier 3 workers spawned by the MMA, the ConductorEngine or caller is responsible for constructing the engine (typically with the same configuration as the main discussion). - If a project disables RAG, `rag_engine=None` is passed to `send()` and the integration is a no-op. **Why caller-owned?** The RAG engine is decoupled from `ai_client` so that the same module can be reused by the GUI's RAG panel for direct queries, by MMA workers for ticket-specific retrieval, and by future automation scripts. `ai_client` only knows how to *use* an engine if one is provided. --- ## Components ### `RAGEngine` (`src/rag_engine.py`) The central class. Owns the embedding provider and the vector store, exposes high-level methods for indexing and search. ```python class RAGEngine: def __init__(self, config: models.RAGConfig, base_dir: str = "."): ... ``` **Construction**: Takes a `RAGConfig` (from `src/models.py`) and a `base_dir`. The config specifies the embedding provider type, the vector store path, the chunk size, and the chunk overlap. **Internal state**: - `embedding_provider: BaseEmbeddingProvider` — set by `_init_embedding_provider` - `client: chromadb.PersistentClient` — the chroma client (or the string `"mock"` in mock mode) - `collection: chromadb.Collection` — the actual collection (or `"mock"` in mock mode) - `chunk_size: int` — character count per chunk - `chunk_overlap: int` — overlap between adjacent chunks ### Embedding Providers Two providers are implemented; new ones can be added by subclassing `BaseEmbeddingProvider`. #### `BaseEmbeddingProvider` ```python class BaseEmbeddingProvider: def embed(self, texts: List[str]) -> List[List[float]]: """Embed a batch of texts. Returns one vector per input text.""" ... ``` A contract: `embed()` takes a list of strings and returns a list of equal-length float vectors. The vector dimensionality is provider-specific (e.g., 384 for `all-MiniLM-L6-v2`, 768 for `gemini-embedding-001`). #### `LocalEmbeddingProvider` Uses **sentence-transformers** (`all-MiniLM-L6-v2` by default) for embedding. - **Pros**: Fully local, no API quota, deterministic. - **Cons**: Lower-quality embeddings than cloud models for code; CPU/GPU usage during indexing. - **Default model**: `all-MiniLM-L6-v2` (384 dimensions, ~80MB download on first use). ```python class LocalEmbeddingProvider(BaseEmbeddingProvider): def __init__(self, model_name: str = 'all-MiniLM-L6-v2'): ... ``` #### `GeminiEmbeddingProvider` Uses the **Gemini Embedding 001** model via the google-genai SDK. - **Pros**: Higher-quality embeddings, especially for code; no local model download. - **Cons**: Requires Gemini API key, network round-trip per embedding call, subject to API quotas. ```python class GeminiEmbeddingProvider(BaseEmbeddingProvider): def __init__(self, model_name: str = 'gemini-embedding-001'): ... ``` #### Lazy Loading The heavy dependencies (`sentence_transformers`, `google.genai`, `chromadb`) are loaded lazily via `_get_sentence_transformers()`, `_get_google_genai()`, `_get_chromadb()`. This means RAG is opt-in: a project that doesn't enable RAG pays no import-time cost. ### Vector Store ChromaDB is the default persistent vector store. The store is created at `/.slop_cache/chroma_/` (auto-generated from `VectorStoreConfig.collection_name`, default `"manual_slop"`). The `.slop_cache` location is intentional — it co-locates the chroma index with the existing per-project cache layout. ```python def _init_vector_store(self): vs_config = self.config.vector_store if vs_config.provider == 'chroma': db_path = os.path.abspath(os.path.join( self.base_dir, ".slop_cache", f"chroma_{vs_config.collection_name}" )) os.makedirs(db_path, exist_ok=True) chromadb, Settings = _get_chromadb() self.client = chromadb.PersistentClient(path=db_path) self.collection = self.client.get_or_create_collection(name=vs_config.collection_name) self._validate_collection_dim() elif vs_config.provider == 'mock': self.client = "mock" self.collection = "mock" else: raise ValueError(f"Unknown vector store provider: {vs_config.provider}") ``` **Backends** (`VectorStoreConfig.provider`): - `chroma` (default for real use) — local persistent, single-process - `mock` — no-op collection (for tests / RAG-disabled paths) The `mcp_server` + `mcp_tool` fields in `VectorStoreConfig` are placeholders for the future External RAG Bridge via MCP (e.g., a remote vector database server); not yet implemented. ### Chunking Strategies Two strategies are implemented. The choice is made per-file based on extension and config. #### Character-Based (`_chunk_text`) Default for non-Python files and for Python files when AST chunking is disabled. ```python def _chunk_text(self, content: str) -> List[str]: """Character-based chunking with overlap.""" chunks = [] start = 0 while start < len(content): end = min(start + self.chunk_size, len(content)) chunks.append(content[start:end]) if end >= len(content): break start = end - self.chunk_overlap return chunks ``` - **Default chunk size**: 1000 characters - **Default overlap**: 200 characters - **Edge cases**: Empty files return `[]`; single-chunk files return `[content]`. #### AST-Aware (`_chunk_code`) Used for `.py` files when `RAGConfig.ast_chunking_enabled = True`. ```python def _chunk_code(self, content: str, file_path: str) -> List[str]: """AST-aware chunking for Python code.""" # Parses with stdlib ast # Splits on top-level def/class boundaries # Each chunk is a complete top-level definition with its docstring ... ``` - **Strategy**: Each top-level function, class, or constant block becomes one chunk. Docstrings are preserved as the first line of the chunk for context. - **Pros**: Semantic boundaries produce more meaningful retrieval results. A query for "how does X work" is more likely to return the entire definition of X rather than a fragment. - **Cons**: Requires valid Python; syntax errors fall back to character-based chunking. The chunker uses stdlib `ast` (not tree-sitter) to avoid pulling tree-sitter for a feature that only handles Python. --- ## Data Flow ### Indexing Flow When a project is loaded with RAG enabled, the `RAGEngine` is populated by indexing all tracked files. ``` 1. Project load: AppController reads [rag] section from manual_slop.toml 2. AppController constructs RAGEngine(config) 3. RAGEngine._init_vector_store() creates/loads ChromaDB collection - Calls _validate_collection_dim() to detect/recover from dim mismatch 4. For each tracked file (parallelized): a. Read content b. Choose chunker based on extension and config c. For each chunk: call embedding_provider.embed([chunk]) d. Add to vector store with metadata {path, chunk_index, ...} 5. Indexing complete; engine is ready for queries ``` **Parallelization**: The indexing pipeline uses `ThreadPoolExecutor` for parallel embedding calls (the embedding step is the bottleneck). The chunking is fast and sequential per file. **Incremental Updates**: When a file's `mtime` changes (detected by `pathlib.Path.stat().st_mtime`), `delete_documents_by_path()` is called first, then the file is re-indexed. This is critical for the auto-sync flow (see Configuration below). **Path resolution resilience**: `index_file()` falls back to `os.getcwd()` if the `base_dir`-relative path doesn't exist. This handles batched test conditions where the subprocess CWD differs from the project root (e.g., a test chdir'ing into `tests/artifacts/live_gui_workspace_*/` for fixture isolation). Without the fallback, indexing silently skipped files in those conditions. ### Dimension Mismatch Protection `_init_vector_store()` calls `_validate_collection_dim()` after creating the collection. The validation inspects the first existing vector's dim and compares it to the current embedding provider's output. On mismatch (e.g., the user switched from Gemini 3072-dim to local 384-dim, or vice versa, or a prior run populated the collection with a different model), the chroma directory is wiped via `shutil.rmtree` (with the client closed first to release file handles) and the collection is recreated with the correct dim. **Why this exists:** Without validation, dim-mismatched upserts silently corrupt the collection. The next `search()` raises `chromadb.errors.InvalidDimensionError: Collection expecting embedding with dimension of X, got Y`, the AI request never reaches `'done'` status, and the live_gui test polls timeout at 50×0.5s = 25s. This pattern was the dominant cause of `tier-3-live_gui` failures in the 2026-06-08 to 2026-06-10 window. Regression tests in `tests/test_rag_engine.py`: `test_rag_collection_dim_mismatch_recreates_collection`, `test_rag_collection_dim_match_preserves_collection`. ### Query Flow When `ai_client.send(rag_engine=engine)` is called: ``` 1. send() receives user_message 2. If rag_engine is not None: a. rag_engine.search(user_message, top_k=5) -> list of {text, metadata, distance} b. If results non-empty: inject [RETRIEVED CONTEXT] block into md_content c. The block contains the top_k fragments, formatted as: ``` [RETRIEVED CONTEXT] File: path/to/file.py (chunk 0) File: path/to/another.py (chunk 2) ... ``` 3. send() proceeds to the provider call with the augmented md_content ``` The injection point is **before** the system prompt construction. This means the retrieved context is treated as part of the project's tracked content, not as ad-hoc advice. ### Public Methods ```python # Index a single file rag_engine.index_file(path: str) -> None # Search the index rag_engine.search(query: str, top_k: int = 5) -> List[Dict[str, Any]] # Returns: [{"text": str, "metadata": dict, "distance": float}, ...] # Index management rag_engine.add_documents(ids: List[str], texts: List[str], metadatas: Optional[List[dict]] = None) -> None rag_engine.delete_documents(ids: List[str]) -> None rag_engine.delete_documents_by_path(path: str) -> None rag_engine.get_all_indexed_paths() -> List[str] rag_engine.is_empty() -> bool ``` --- ## Configuration RAG is configured via the project's `manual_slop.toml`: ```toml [rag] enabled = true embedding_provider = "gemini" # or "local" [rag.vector_store] provider = "chroma" # "chroma" | "mock" collection_name = "manual_slop" # the chroma subdir under .slop_cache/ url = "" # future: external HTTP vector store api_key = "" # future: external HTTP auth mcp_server = "" # future: MCP-based external RAG bridge mcp_tool = "" # future: tool name on the MCP server [rag] chunk_size = 1000 chunk_overlap = 200 ``` ### `RAGConfig` + `VectorStoreConfig` Schema (`src/models.py`) ```python @dataclass class VectorStoreConfig: provider: str # "chroma" | "mock" url: Optional[str] = None # future: external HTTP api_key: Optional[str] = None # future: external HTTP auth collection_name: str = "manual_slop" mcp_server: Optional[str] = None # future: MCP bridge mcp_tool: Optional[str] = None # future: MCP tool name @dataclass class RAGConfig: enabled: bool = False vector_store: VectorStoreConfig = field(default_factory=lambda: VectorStoreConfig(provider='mock')) embedding_provider: str = 'gemini' # "gemini" | "local" chunk_size: int = 1000 chunk_overlap: int = 200 ``` > **What about the fields the old doc showed?** The 2026-06-10 docs sync verified against `src/models.py:1029-1040` that the previous `RAGConfig` schema was **stale** (predated the schema refactor) — most of the fields it listed did not exist in the real dataclass. Specifically: `ast_chunking_enabled` does not exist anywhere in `src/` (there is no `ChunkingConfig` class — I claimed one existed in an earlier draft of this note and was wrong; flagging the correction here); `vector_store_backend` and `vector_store_path` never existed on `RAGConfig` (they were a flattened version of the now-nested `VectorStoreConfig`); `auto_index_on_load` and `auto_sync_interval_seconds` do not exist anywhere in `src/` (they were aspirational; the actual index-on-load and auto-sync behavior is wired in `RAGEngine` and the controller's `mma_state_update` flow, not via persisted config); `top_k` IS a real thing but it is a **runtime parameter** to `RAGEngine.search(query, top_k=5)` and `RAGEngine._search_mcp(query, top_k=5)` (`src/rag_engine.py:339, 322`), not a field on `RAGConfig` — the old doc confused "config field" with "search parameter." ### Behavior When Disabled If `enabled = false` (the default), `RAGEngine` is never constructed. `ai_client.send()` receives `rag_engine=None` and the integration is a no-op. The lazy-loading of `chromadb`, `sentence_transformers`, and `google.genai` is also skipped, so there is zero overhead for projects that don't use RAG. ### Auto-Sync When `auto_sync_interval_seconds > 0`, a background thread periodically scans tracked files for `mtime` changes and re-indexes them. This keeps the vector store consistent with on-disk changes without requiring explicit user action. The sync uses `pathlib.Path.stat().st_mtime` for change detection (same mechanism as the file cache in `file_cache.py`). For very large projects, the sync can be tuned to skip files above a size threshold. --- ## Cross-System Integration ### `ai_client.send()` Integration See [guide_architecture.md#rag-integration](guide_architecture.md#rag-integration) for the full dispatch flow. Summary: ```python def send(md_content, user_message, ..., rag_engine=None) -> str: if rag_engine is not None: retrieved = rag_engine.search(user_message, top_k=rag_engine.config.top_k) if retrieved: md_content = _inject_rag_context(md_content, retrieved) ... ``` The injection is a no-op if: - `rag_engine is None` - `rag_engine.is_empty()` (index has no documents) - `search()` returns no results above the distance threshold ### MMA Worker Integration The ConductorEngine does not construct `RAGEngine` itself. Workers receive context via `md_content` which is built by the caller. To use RAG in workers: 1. Construct a `RAGEngine` in the caller (typically `AppController` or test harness). 2. Pass it to `multi_agent_conductor.run_worker_lifecycle(..., rag_engine=...)` (if supported) or to the test invocation. 3. The worker passes it to `ai_client.send(rag_engine=...)`. Note: As of 2026-06-02, the direct `rag_engine` parameter on `run_worker_lifecycle` is **not yet implemented**. Workers currently rely on the `md_content` already being augmented by the caller, or on Tier 4 / Tier 2 setting up the augmentation before spawning workers. ### GUI Integration The GUI's RAG panel (under AI Settings → RAG) provides: - **Status indicator** — `RAGEngine.is_empty()` → "Empty" / "Indexed N chunks" - **Manual search box** — for testing retrieval quality without sending a full AI call - **Re-index button** — forces a full rebuild of the index - **Settings editor** — modifies `RAGConfig` fields and writes back to `manual_slop.toml` The RAG panel also surfaces the **auto-sync status** (last sync time, files indexed, files pending re-index). --- ## Testing ### Unit Tests - `tests/test_rag_engine.py` — `RAGEngine` basic lifecycle with mock ChromaDB and mock embedding provider - `tests/test_rag_integration.py` — End-to-end indexing + search + retrieval ### Simulation Tests - `tests/test_rag_gui_presence.py` — Verifies the RAG panel renders correctly - `tests/test_rag_visual_sim.py` — Visual verification of the RAG search results panel ### Stress Tests - `tests/test_rag_phase4_stress.py` — Indexes 1000+ files, measures retrieval latency - `tests/test_rag_phase4_final_verify.py` — End-to-end verification of RAG-augmented AI responses ### Test Patterns The standard pattern for testing RAG-augmented calls: ```python def test_rag_augmented_send(live_gui): # 1. Set up project with RAG enabled client.set_rag_config(enabled=True, embedding_provider="local") client.reindex_project() # 2. Send a question that requires retrieval response = client.send("How does the Execution Clutch work?") # 3. Verify the response references the retrieved content # (The exact assertion depends on what was indexed) assert response ``` For unit tests that don't need real embedding models, the `BaseEmbeddingProvider` is mocked to return deterministic vectors (e.g., based on the hash of the input text). --- ## Edge Cases & Limitations 1. **Empty Index**: If the index has no documents, `search()` returns `[]` and no context is injected. The AI call proceeds normally with just the explicit file context. 2. **Network Failures (Gemini Embeddings)**: If the Gemini API is unreachable, `GeminiEmbeddingProvider.embed()` raises an exception. The caller (typically `_chunk_code` → `index_file` → RAG indexer) should handle this gracefully and either retry or fall back to the local provider. 3. **Stale Index**: Auto-sync runs periodically but not on every read. If a file is changed between sync intervals, the index may be stale. The `delete_documents_by_path` + `index_file` cycle is atomic per file, so a partial sync leaves the index in a consistent (if incomplete) state. 4. **Large Files**: A single file larger than `chunk_size` is split into multiple chunks with overlap. There's no upper limit on the number of chunks per file, but very large files (>10MB) may slow down indexing significantly. 5. **Binary Files**: RAG only handles text files. Binary files (images, compiled Python, etc.) are skipped during indexing with a warning logged to `comms_log`. 6. **Cross-Project Queries**: The vector store is per-project (`/.rag/chroma/`). Cross-project retrieval is **not** supported; each project has its own isolated index. 7. **Concurrent Writes**: ChromaDB's PersistentClient is single-writer. If multiple processes try to write to the same index simultaneously, ChromaDB will raise. Manual Slop uses a `threading.Lock` to serialize writes from the auto-sync thread and the manual re-index button. --- ## Future Work - **External RAG Bridge** — Connect to remote vector databases (e.g., a managed Pinecone or Weaviate) via MCP. The `_search_mcp` method (`src/rag_engine.py:322`) IS a real implementation: when `RAGConfig.vector_store.provider == "mcp"`, `RAGEngine.search()` dispatches to `_search_mcp()` which calls `mcp_client.async_dispatch("rag_search", {"query": ..., "top_k": ...})`. The MCP-bridge config (`mcp_server`, `mcp_tool`) lives on `VectorStoreConfig`. The bridge wires up the rest of the RAG pipeline to a remote vector store; no per-vendor `_init_vector_store` branch is needed because the MCP server owns that. - **Hybrid Search** — Combine dense (vector) retrieval with sparse (BM25) retrieval for better recall on code keywords. - **Re-ranking** — Apply a cross-encoder reranker to the top-k results before injection to improve precision. - **Caching** — Cache query results in memory to avoid re-embedding for repeated questions. - **Provider Routing** — Allow per-query provider selection (e.g., use Gemini for general queries, local for code). See [guide_tools.md](guide_tools.md) for the MCP tool inventory; see [guide_architecture.md](guide_architecture.md) for the dispatch pipeline.