Skip to main content
LLM Checker uses a deterministic pipeline so that the same inputs always produce the same ranked output, with explicit policy outcomes for governance workflows.

​
Pipeline Flowchart

​
Component Responsibilities

LayerResponsibility
Input layerCollects runtime constraints from hardware detection, local inventory, dynamic registry data, and CLI flags
Normalization layerDeduplicates identifiers/tags and builds a canonical candidate set
Selection layerFilters by use case, selects the best fitting quantization, and computes deterministic Q/S/F/C scores
Governance layerApplies policy rules in audit or enforce mode and records explicit violation metadata
Output layerReturns ranked recommendations plus machine-readable compliance artifacts when requested

​
Execution Stages

The pipeline runs through six sequential stages:
  1. Hardware profiling — Detect CPU/GPU/RAM and effective backend capabilities.
  2. Model pool assembly — Merge dynamic scraped catalog (or curated fallback) with locally installed models.
  3. Candidate filtering — Keep only relevant models for the requested use case.
  4. Fit selection — Choose the best quantization for the available memory budget.
  5. Deterministic scoring — Score each candidate across Quality, Speed, Fit, and Context.
  6. Policy + ranking — Apply optional policy checks, then rank and return actionable commands.

​
Project Structure

src/
  models/
    deterministic-selector.js  # Primary selection algorithm
    scoring-config.js          # Centralized scoring weights
    scoring-engine.js          # Advanced scoring (smart-recommend)
    catalog.json               # Curated fallback catalog (35+ models, only if dynamic pool unavailable)
  ai/
    multi-objective-selector.js  # Multi-objective optimization
    ai-check-selector.js         # LLM-based evaluation
  hardware/
    detector.js                # Hardware detection
    unified-detector.js        # Cross-platform detection
  data/
    model-database.js          # SQLite storage (optional)
    sync-manager.js            # Database sync from Ollama registry
bin/
  enhanced_cli.js              # CLI entry point

​
Key Source Files

The primary selection algorithm. Implements the 4D scoring pipeline (Q/S/F/C), quantization fit selection, MoE memory estimation, and runtime-aware MoE speed overhead. Used by check and recommend.
Centralized store for all scoring weight configurations. Exports three weight sets:
  • DETERMINISTIC_WEIGHTS — per-category [Q, S, F, C] arrays for the primary selector
  • MULTI_OBJECTIVE_WEIGHTS — five-factor weights including hardwareMatch
  • SCORING_ENGINE_WEIGHTS — {Q, S, F, C} objects with additional presets for smart-recommend and search
Advanced scoring engine that powers smart-recommend and search. Reads weight presets from scoring-config.js and applies runtime-aware MoE speed estimation from moe-assumptions.js.
Built-in curated fallback catalog of 35+ models. Only loaded when the dynamic scraped Ollama pool is unavailable. Covers major families: Qwen, Llama, DeepSeek, Phi, Gemma, Mistral, CodeLlama, LLaVA, and embeddings models.
Platform-specific hardware detection logic. Reads CPU, RAM, and GPU information. Handles Apple Silicon unified memory, NVIDIA CUDA, AMD ROCm VRAM normalization, and Intel Arc detection.
Cross-platform detection layer that merges outputs from platform-specific detectors. Preserves dedicated and integrated GPU topology separately for hybrid systems. Produces the canonical hardware profile consumed by the selection pipeline.
Optional SQLite storage layer (requires sql.js). Persists the scraped Ollama catalog to disk and serves queries for search and smart-recommend.
Fetches the latest model catalog from the Ollama registry and writes it to the local SQLite database. Invoked by llm-checker sync.
The CLI entry point. Parses arguments, routes commands, enforces policy precedence (--policy > --calibrated > deterministic fallback), and formats output for both human-readable terminal display and machine-readable JSON.

Build docs developers (and LLMs) love

Get started for freeTalk to us