llm-checker 3.5.14 → 3.6.1

package/README.md

@@ -573,6 +573,19 @@ This makes integrated GPUs visible even when the selected runtime backend is sti
 llm-checker recommend
 ```
 
+As of the scoring unification (#96), `check`, `recommend`, and `smart-recommend`
+all derive their ranking from **one canonical scoring core**
+(`DeterministicModelSelector` via `src/models/scoring-core.js`), so identical
+`(model, hardware)` inputs score identically across all three and the
+high-capacity right-sizing floor applies everywhere. They differ only in their
+model **source** and **presentation**, not in how a given model is ranked:
+
+| Command | Role | Ranking core |
+|---------|------|--------------|
+| `recommend` | Canonical model recommendations by category | Shared core (reference output) |
+| `check` | Full hardware-compatibility report with a recommendation card | Shared core (consistent ranking, fit-oriented report) |
+| `smart-recommend` | Catalog/DB-backed recommendations with a detailed score breakdown | Shared core (same ordering + scores) |
+
 Use optimization profiles to steer ranking by intent:
 
 ```bash
@@ -695,7 +708,7 @@ Three scoring systems are available, each optimized for different workflows:
 | `reasoning` | 60% | 10% | 20% | 10% |
 | `multimodal` | 50% | 15% | 20% | 15% |
 
-**Scoring Engine** (used by `smart-recommend` and `search`):
+**Scoring Engine** (used by `search` for catalog scoring; `smart-recommend`'s final ranking is produced by the shared scoring core — see #96):
 
 | Use Case | Quality | Speed | Fit | Context |
 |----------|:-------:|:-----:|:---:|:-------:|

package/analyzer/compatibility.js

@@ -428,6 +428,11 @@ class CompatibilityAnalyzer {
     }
 
     parseModelSize(sizeString) {
+        // Guard non-string input (undefined / a numeric size) — this runs for every
+        // model in calculateModelCompatibility, so one bad entry must not crash the
+        // whole analysis. Matches the guard in analyzer/performance.js.
+        if (typeof sizeString !== 'string' || !sizeString.trim()) return 1;
+
         const match = sizeString.match(/(\d+\.?\d*)[BM]/i);
         if (!match) return 1;
 

package/analyzer/performance.js

@@ -363,12 +363,13 @@ class PerformanceAnalyzer {
     }
 
     estimateLoadTime(model, hardware) {
+        // ~2 GB per 1B params (fp16-ish) on-disk approximation.
         const modelSizeGB = this.parseModelSize(model.size) * 2;
 
-        let loadTimeSeconds = modelSizeGB * 2;
-
-
-        loadTimeSeconds *= 0.7;
+        // Fold the previous `* 2` then `* 0.7` two-step (a leftover from an
+        // incomplete edit, with a dead blank line) into one documented factor:
+        // ~1.4 s of load time per GB before hardware adjustments.
+        let loadTimeSeconds = modelSizeGB * 1.4;
 
         const cpuSpeedFactor = Math.max(0.5, Math.min(1.5, (hardware.cpu.speed || 2.5) / 2.5));
         loadTimeSeconds /= cpuSpeedFactor;

package/bin/cli.js

@@ -11,44 +11,10 @@ if (!Number.isFinite(majorNodeVersion) || majorNodeVersion < 16) {
     process.exit(1);
 }
 
-function preprocessAiCheckModelsArg(argv) {
-    const normalizedArgs = [];
-    let modelsFilter = null;
-    let sawAiCheck = false;
-
-    for (let index = 0; index < argv.length; index += 1) {
-        const token = argv[index];
-
-        if (token === 'ai-check') {
-            sawAiCheck = true;
-        }
-
-        if (sawAiCheck && token === '--models') {
-            const nextToken = argv[index + 1];
-            if (nextToken && !nextToken.startsWith('-')) {
-                modelsFilter = nextToken;
-                index += 1;
-            }
-            continue;
-        }
-
-        if (sawAiCheck && token.startsWith('--models=')) {
-            modelsFilter = token.slice('--models='.length);
-            continue;
-        }
-
-        normalizedArgs.push(token);
-    }
-
-    return { args: normalizedArgs, modelsFilter };
-}
-
-const preprocessedArgs = preprocessAiCheckModelsArg(process.argv.slice(2));
-
-if (typeof preprocessedArgs.modelsFilter === 'string' && preprocessedArgs.modelsFilter.trim()) {
-    process.env.LLM_CHECKER_AI_CHECK_MODELS = preprocessedArgs.modelsFilter.trim();
-}
-
-process.argv = [process.argv[0], process.argv[1], ...preprocessedArgs.args];
+// `ai-check --models <list>` is now a real commander option handled in
+// enhanced_cli.js (and the AICheckSelector applies it as a candidate filter), so
+// the previous argv-rewriting shim — which stripped the flag and stashed it in an
+// env var that nothing read — is gone. LLM_CHECKER_AI_CHECK_MODELS still works as
+// an explicit fallback for the same filter.
 
 require('./enhanced_cli');

package/bin/enhanced_cli.js

@@ -58,7 +58,7 @@ const calibrationManager = new CalibrationManager();
 
 const COMMAND_HEADER_LABELS = {
     'hw-detect': 'Hardware Detection',
-    'smart-recommend': 'Smart Recommend',
+    'smart-recommend': 'Smart Recommend (Experimental)',
     search: 'Model Search',
     sync: 'Database Sync',
     'mcp-setup': 'Claude MCP Setup',
@@ -79,6 +79,19 @@ function showAsciiArt(command) {
     renderCommandHeader(label);
 }
 
+const RECOMMENDATION_COMMAND_NOTES = {
+    check: 'Compatibility report: shows hardware fit first. Use `llm-checker recommend` for canonical ranked model picks.',
+    recommend: 'Canonical recommendations: deterministic hardware-aware selector by category.',
+    'smart-recommend': 'Experimental scoring engine: results can differ from `recommend` while this path is being unified.'
+};
+
+function displayRecommendationCommandNote(command) {
+    const note = RECOMMENDATION_COMMAND_NOTES[command];
+    if (!note) return;
+    console.log(chalk.gray(`Recommendation mode: ${note}`));
+    console.log('');
+}
+
 // Function to search Ollama models by use case
 function getOllamaCacheFile(filename) {
     try {
@@ -3295,6 +3308,7 @@ Policy scope:
     )
     .action(async (options) => {
         showAsciiArt('check');
+        displayRecommendationCommandNote('check');
         try {
             // Use verbose progress unless explicitly disabled
             const verboseEnabled = options.verbose !== false;
@@ -3462,11 +3476,8 @@ Policy scope:
 
 program
     .command('ollama')
-    .description('Manage Ollama integration with hardware compatibility')
-    .option('-l, --list', 'List installed models with compatibility scores')
-    .option('-r, --running', 'Show running models with performance data')
-    .option('-c, --compatible', 'Show only hardware-compatible installed models')
-    .option('--recommendations', 'Show installation recommendations')
+    .description('Check Ollama integration status (use `installed` to rank installed models)')
+    .option('-l, --list', 'List installed models ranked by compatibility (runs `installed`)')
     .action(async (options) => {
         showAsciiArt('ollama');
         const spinner = ora('Checking Ollama integration...').start();
@@ -3492,7 +3503,9 @@ program
             spinner.succeed(`Ollama integration active`);
 
             if (options.list) {
-                console.log('Ollama models list feature coming soon...');
+                console.log(chalk.cyan('\nRun `llm-checker installed` to rank your installed models by compatibility and use-case.'));
+            } else {
+                console.log(chalk.gray('\nTip: `llm-checker installed` ranks installed models; `llm-checker recommend` suggests models for your hardware.'));
             }
 
         } catch (error) {
@@ -3520,10 +3533,18 @@ program
             const availability = await ollamaClient.checkOllamaAvailability();
             if (!availability.available) {
                 spinner.fail('Ollama not available');
-                console.log(chalk.red('\n' + availability.error));
-                if (availability.hint) {
-                    console.log(chalk.yellow('Hint: ' + availability.hint));
+                if (options.json) {
+                    // --json must always emit parseable JSON on stdout (the success
+                    // path prints an array); previously these branches printed
+                    // ANSI-colored prose and broke `installed --json | jq`.
+                    console.log(JSON.stringify([], null, 2));
+                } else {
+                    console.log(chalk.red('\n' + availability.error));
+                    if (availability.hint) {
+                        console.log(chalk.yellow('Hint: ' + availability.hint));
+                    }
                 }
+                process.exitCode = 1;
                 return;
             }
 
@@ -3531,8 +3552,12 @@ program
             const installedModels = await ollamaClient.getLocalModels();
             if (!installedModels || installedModels.length === 0) {
                 spinner.fail('No models installed');
-                console.log(chalk.yellow('\nNo Ollama models found. Install one with:'));
-                console.log(chalk.cyan('  ollama pull llama3.2:3b'));
+                if (options.json) {
+                    console.log(JSON.stringify([], null, 2));
+                } else {
+                    console.log(chalk.yellow('\nNo Ollama models found. Install one with:'));
+                    console.log(chalk.cyan('  ollama pull llama3.2:3b'));
+                }
                 return;
             }
 
@@ -3851,6 +3876,7 @@ Calibrated routing examples:
     )
     .action(async (options) => {
         showAsciiArt('recommend');
+        displayRecommendationCommandNote('recommend');
         try {
             const verboseEnabled = options.verbose !== false;
             const checker = new (getLLMChecker())({ verbose: verboseEnabled });
@@ -4365,6 +4391,7 @@ program
     .option('--ctx <number>', 'Target context length', '8192')
     .option('-e, --evaluator <model>', 'Evaluator model (auto for best available)', 'auto')
     .option('-w, --weight <number>', 'AI weight (0.0-1.0, default 0.3)', '0.3')
+    .option('-m, --models <list>', 'Restrict evaluation to these models (comma-separated)')
     .action(async (options) => {
         showAsciiArt('ai-check');
         // Check if Ollama is installed first
@@ -4377,14 +4404,33 @@ program
             
             const aiCheckSelector = new AICheckSelector();
             
+            // Validate numeric options up front: bad input (e.g. --weight abc, --top
+            // foo) used to flow through as NaN, which survived the downstream clamp
+            // and made the displayed AI weight and every weighted score NaN.
+            const weight = Number.parseFloat(options.weight);
+            const top = Number.parseInt(options.top, 10);
+            const ctx = options.ctx ? Number.parseInt(options.ctx, 10) : undefined;
+            const invalidNumeric =
+                (!Number.isFinite(weight) || weight < 0 || weight > 1) ? `--weight ${options.weight} (use a number from 0.0 to 1.0)`
+                : (!Number.isInteger(top) || top <= 0) ? `--top ${options.top} (use a positive integer)`
+                : (ctx !== undefined && (!Number.isInteger(ctx) || ctx <= 0)) ? `--ctx ${options.ctx} (use a positive integer)`
+                : null;
+            if (invalidNumeric) {
+                spinner.stop();
+                console.error(chalk.red(`Invalid ${invalidNumeric}`));
+                process.exitCode = 1;
+                return;
+            }
+
             const checkOptions = {
                 category: options.category,
-                top: parseInt(options.top),
-                ctx: options.ctx ? parseInt(options.ctx) : undefined,
+                top,
+                ctx,
                 evaluator: options.evaluator,
-                weight: parseFloat(options.weight)
+                weight,
+                models: options.models || process.env.LLM_CHECKER_AI_CHECK_MODELS || undefined
             };
-            
+
             spinner.stop();
             
             const result = await aiCheckSelector.aiCheck(checkOptions);
@@ -4731,8 +4777,20 @@ program
             });
 
             if (searchResults.length === 0) {
-                if (spinner) spinner.info('No models found matching your query');
                 syncManager.close();
+                if (options.json) {
+                    // --json must always emit parseable JSON, even on zero matches
+                    // (previously it printed nothing and broke `search ... --json | jq`).
+                    console.log(JSON.stringify({
+                        query,
+                        all: [],
+                        recommendations: [],
+                        insights: [],
+                        meta: { afterFiltering: 0, totalMatches: 0 }
+                    }, null, 2));
+                } else if (spinner) {
+                    spinner.info('No models found matching your query');
+                }
                 return;
             }
 
@@ -4806,7 +4864,7 @@ program
 
 program
     .command('smart-recommend')
-    .description('Get intelligent model recommendations using the new scoring engine')
+    .description('Experimental recommendations using the alternate scoring engine')
     .option('-u, --use-case <case>', 'Optimize for use case', 'general')
     .option('-l, --limit <n>', 'Maximum number of recommendations', '5')
     .option('--target-tps <n>', 'Target tokens per second', '20')
@@ -4814,8 +4872,19 @@ program
     .option('--include-vision', 'Include vision/multimodal models')
     .option('--include-embeddings', 'Include embedding models')
     .option('-j, --json', 'Output as JSON')
+    .addHelpText(
+        'after',
+        `
+Recommendation engine note:
+  smart-recommend is experimental and may intentionally differ from recommend.
+  Use "llm-checker recommend" for canonical package recommendations.
+`
+    )
     .action(async (options) => {
-        if (!options.json) showAsciiArt('smart-recommend');
+        if (!options.json) {
+            showAsciiArt('smart-recommend');
+            displayRecommendationCommandNote('smart-recommend');
+        }
         const SyncManager = require('../src/data/sync-manager');
         const IntelligentSelector = require('../src/models/intelligent-selector');
         const UnifiedDetector = require('../src/hardware/unified-detector');