SLM RAG
Local Retrieval-Augmented Generation engine for offline document indexing, dense vector search, and grounded Q&A.
01 / Overview & Capabilities
Local Retrieval-Augmented Generation engine for offline document indexing, dense vector search, and grounded Q&A.
Key Features
- Offline 1024-d (mxbai-embed-large) & 384-d CPU embedding generation + cross-encoder reranking via the 10-Plane Harness
- Local vector similarity indexing
- Grounded factual document question answering
- Strict context window enforcement
02 / Standalone OS Harness Installers, CLI & Python SDK
In SLMAgents 2.0, SLM RAG executes inside the 10-Plane Pure-CPU Agent Harness (slm_harness). Download a self-extracting Individual Agent + Harness OS Installer for macOS, Windows, or Linux (bundling slm_rag + slm_harness + the ./slm CLI), install the full 26-agent desktop suite, or install via Python pip:
1. Individual SLM RAG + 10-Plane Harness Installers (All OSes)
slm_rag + slm_harness + ./slm CLIslm_rag + slm_harnessslm_rag + slm_harness + ./slm CLI2. Full 26-Agent Studio & Desktop OS Installers
3. 10-Plane Harness CLI (./slm) & Python pip Installation
# Option A: Install via PyPI or editable source with the 10-Plane Harness $ pip install slm-rag $ pip install -e ./slm_harness -e ./slm_rag # Option B: Run directly through the Unified 10-Plane Pure-CPU Harness CLI (./slm) $ ./slm ask "What is the total Q3 net revenue in @financial_report.pdf?" $ ./slm chat # Interactive Rich TUI with 18 slash commands (/trace, /models, /arena, /undo) $ ./slm serve # 100% Offline Web AI Studio + OpenAI /v1 API at http://127.0.0.1:7860
03 / 10-Plane Pure-CPU Harness Integration (slm_harness)
Rather than executing as an isolated prompt wrapper, SLM RAG is natively orchestrated by the SLMAgents 2.0 10-Plane Pure-CPU Harness (via slm_harness.runtime.legacy_bridge and CascadeRouter). Every call benefits from hybrid llama.cpp GGUF (Q4_K_M) + Quantized INT4/INT8 CPU execution, watermarked LRU memory governance, grammar-constrained decoding, and deterministic verification at $0 API cost:
| Harness Plane | Subsystem | How It Powers SLM RAG on Pure CPU |
|---|---|---|
| Plane 6 | CascadeRouter (<35ms) |
Stage 2 (Document/Query Signature) & Stage 3 (1024-d Embedding Margin) |
| Plane 2 & 5 | MemoryArena & Hybrid CPU Runtime |
Specialist Role: embeddings (mxbai-embed-large INT8) + reranker (bge-reranker-base) + primary (Qwen3-4B GGUF Q4_K_M / Llama-3.2-3B INT4). Managed inside Watermark LRU MemoryArena (lite 2.6 GB / balanced 4.7 GB / power 8.2 GB) with zero-copy session sharing and automatic OOM tier step-down (8B -> 4B -> 3B -> 0.8B). |
| Plane 7 & 8 | Deterministic Verifier & Governance | Plane 7 Citation Grounding Verifier ([1], [2] span-level verification against retrieved chunks) + Context Distillation (Top-K=2) |
| Plane 1, 3 & 4 | Kernel, Artifact Scratchpad & 4-Tier Memory | SQLite WAL CheckpointStore (/fork, /rewind, /undo), content-addressed artifact://<sha12> handles to prevent prompt bloat, and hybrid SQLite FTS5 BM25 + HNSW vector memory. |
04 / Source Checkout (Git Sparse-Checkout)
Clone this agent alongside the 10-Plane Pure-CPU Harness (slm_harness) from the monorepo using Git sparse-checkout — no need to download all 26 agents:
Option 1 — Sparse Checkout with 10-Plane Harness (Recommended)
# 1. Create and enter a new directory $ mkdir slm_rag_workspace && cd slm_rag_workspace # 2. Initialise empty git repo and add remote $ git init $ git remote add origin https://github.com/t00114218-stack/SLMAgents.git # 3. Enable sparse-checkout for both slm_harness and slm_rag $ git sparse-checkout init --cone $ git sparse-checkout set slm_harness slm_rag # 4. Pull the 10-Plane Harness and agent source $ git pull origin main
Option 2 — Full Repository Clone
$ git clone https://github.com/t00114218-stack/SLMAgents.git $ cd SLMAgents $ ./slm doctor
Note: After checkout, install the harness and agent in editable mode with pip install -e ./slm_harness -e ./slm_rag or download the pre-packaged OS Harness Installer above.
05 / Configuration & API Reference
Constructor Parameters
Instantiate SLMRag with performance and runtime options:
| Parameter | Type / Default | Description |
|---|---|---|
model_path | str | None | Local GGUF (.gguf) or quantized ONNX model path managed by MemoryArena. Default: None (auto-resolved via slm_harness). |
temperature | float | 0.0 | Sampling temperature. Default: 0.0. |
top_p | float | 0.9 | Nucleus sampling threshold. Default: 0.9. |
max_tokens | int | 256 | Maximum output token limit. Default: 256. |
n_threads | int | 4 | CPU threads. Default: 4. |
Methods
| Method Signature | Return Type | Description |
|---|---|---|
answer(question, chunks, ...) | str | dict | Retrieves matching context chunks and generates factual response. |
Execution Parameters
Complete list of execution parameters accepted by the primary agent method:
| Parameter | Type / Default | Description |
|---|---|---|
question | str | User question string. |
chunks | list[str] | str | Document context text or chunk list. |
instruction | str | None | Extraction instruction guideline. Default: None. |
system_prompt | str | None | System prompt instruction. Default: None. |
user_input | str | None | Contextual key values. Default: None. |
temperature | float | 0.0 | Sampling temperature. Default: 0.0. |
top_p | float | 0.9 | Nucleus sampling probability. Default: 0.9. |
max_tokens | int | 256 | Maximum token limit. Default: 256. |
Quick Start
from slm_rag import SLMRag
rag = SLMRag(temperature=0.0, top_p=0.9, max_tokens=256)
res = rag.answer(
question="What is total Q3 net revenue?",
chunks="Q3 net revenue reached $1.25M., Due date: Sept 2026",
instruction="Extract exact numerical totals",
system_prompt="Factual document extraction",
user_input="Currency: USD",
temperature=0.0,
top_p=0.9,
max_tokens=256
)
print(res)
06 / Verified Execution Logs
Diagnostic execution console output running locally on CPU:
→ INPUT:
Question: What is total Q3 net revenue?
← OUTPUT:
{
'status': '200 OK',
'answer': 'Based on documents: Q3 net revenue is $1.25M USD.',
'retrieved_chunks': 2
}