rag-lite-ts 1.0.2 → 2.0.0

package/dist/core/reranking-strategies.js

@@ -0,0 +1,720 @@
+/**
+ * Cross-Encoder Reranking Strategy for Text Mode
+ *
+ * Adapts the existing CrossEncoderReranker to work with the new RerankingStrategy
+ * interface defined in the Chameleon Multimodal Architecture.
+ */
+import { CrossEncoderReranker } from '../text/reranker.js';
+/**
+ * Cross-Encoder Reranking Strategy Implementation
+ *
+ * Uses the existing CrossEncoderReranker from the text module to provide
+ * reranking functionality for text mode in the new architecture.
+ */
+export class CrossEncoderRerankingStrategy {
+    name = 'cross-encoder';
+    supportedContentTypes = ['text'];
+    isEnabled = true;
+    reranker;
+    modelName;
+    initialized = false;
+    constructor(modelName) {
+        this.modelName = modelName;
+        this.reranker = new CrossEncoderReranker();
+        // Set custom model name if provided
+        if (modelName) {
+            this.reranker.modelName = modelName;
+        }
+    }
+    /**
+     * Initialize the reranker if not already done
+     */
+    async ensureInitialized() {
+        if (!this.initialized) {
+            try {
+                await this.reranker.loadModel();
+                this.initialized = true;
+                this.isEnabled = this.reranker.isLoaded();
+                if (!this.isEnabled) {
+                    console.warn('Cross-encoder reranker failed to load, strategy disabled');
+                }
+            }
+            catch (error) {
+                console.warn(`Cross-encoder reranker initialization failed: ${error instanceof Error ? error.message : 'Unknown error'}`);
+                this.isEnabled = false;
+            }
+        }
+    }
+    /**
+     * Rerank search results using cross-encoder model
+     */
+    rerank = async (query, results, contentType) => {
+        // If strategy is disabled, return results unchanged immediately
+        if (!this.isEnabled) {
+            return results;
+        }
+        // Validate content type
+        if (contentType && !this.supportedContentTypes.includes(contentType)) {
+            throw new Error(`Cross-encoder strategy does not support content type '${contentType}'. ` +
+                `Supported types: ${this.supportedContentTypes.join(', ')}`);
+        }
+        // Ensure reranker is initialized
+        await this.ensureInitialized();
+        // If reranker failed to initialize, return results unchanged
+        if (!this.isEnabled) {
+            console.warn('Cross-encoder reranker not enabled, returning results unchanged');
+            return results;
+        }
+        // Filter to only text content if mixed content types are present
+        const textResults = results.filter(result => !result.contentType || result.contentType === 'text');
+        if (textResults.length === 0) {
+            return results; // No text results to rerank
+        }
+        if (textResults.length !== results.length) {
+            console.warn(`Cross-encoder reranker filtering ${results.length - textResults.length} ` +
+                `non-text results from reranking`);
+        }
+        try {
+            // Use the existing reranker implementation
+            const rerankedTextResults = await this.reranker.rerank(query, textResults);
+            // If we filtered results, we need to merge back non-text results
+            if (textResults.length !== results.length) {
+                const nonTextResults = results.filter(result => result.contentType && result.contentType !== 'text');
+                // Append non-text results at the end with their original scores
+                return [...rerankedTextResults, ...nonTextResults];
+            }
+            return rerankedTextResults;
+        }
+        catch (error) {
+            console.warn(`Cross-encoder reranking failed: ${error instanceof Error ? error.message : 'Unknown error'}. ` +
+                `Returning original results.`);
+            return results;
+        }
+    };
+    /**
+     * Configure the reranking strategy
+     */
+    configure(config) {
+        if (config.modelName && typeof config.modelName === 'string') {
+            this.modelName = config.modelName;
+            // Reset initialization to use new model
+            this.initialized = false;
+            this.reranker = new CrossEncoderReranker();
+            this.reranker.modelName = config.modelName;
+        }
+        if (config.enabled !== undefined) {
+            this.isEnabled = Boolean(config.enabled);
+        }
+    }
+    /**
+     * Get metadata about this reranking strategy
+     */
+    getMetadata() {
+        return {
+            description: 'Cross-encoder reranking using transformer models for improved text relevance scoring',
+            requiredModels: [
+                'Xenova/ms-marco-MiniLM-L-6-v2',
+                'cross-encoder/ms-marco-MiniLM-L-6-v2',
+                'cross-encoder/ms-marco-MiniLM-L-2-v2'
+            ],
+            configOptions: {
+                modelName: {
+                    type: 'string',
+                    description: 'Cross-encoder model name to use for reranking',
+                    default: 'Xenova/ms-marco-MiniLM-L-6-v2'
+                },
+                enabled: {
+                    type: 'boolean',
+                    description: 'Enable or disable cross-encoder reranking',
+                    default: true
+                }
+            }
+        };
+    }
+    /**
+     * Check if the strategy is ready to use
+     */
+    async isReady() {
+        await this.ensureInitialized();
+        return this.isEnabled && this.reranker.isLoaded();
+    }
+    /**
+     * Get the current model name being used
+     */
+    getModelName() {
+        return this.reranker.getModelName();
+    }
+    /**
+     * Clean up resources
+     */
+    async cleanup() {
+        // The existing CrossEncoderReranker doesn't have explicit cleanup
+        // but we can reset the initialization state
+        this.initialized = false;
+    }
+}
+/**
+ * Factory function to create a cross-encoder reranking strategy
+ *
+ * This provides a simple way to create the strategy without complex factory patterns,
+ * following the design principle of using simple functions over complex factories.
+ */
+export function createCrossEncoderStrategy(modelName) {
+    return new CrossEncoderRerankingStrategy(modelName);
+}
+/**
+ * Text-Derived Reranking Strategy Implementation
+ *
+ * Converts images to text descriptions using image-to-text models, then applies
+ * cross-encoder reranking to the text descriptions. This enables multimodal
+ * content to be reranked using text-based reranking models.
+ */
+export class TextDerivedRerankingStrategy {
+    name = 'text-derived';
+    supportedContentTypes = ['text', 'image'];
+    isEnabled = true;
+    crossEncoderReranker;
+    imageToTextModel = null;
+    imageToTextModelName = 'Xenova/vit-gpt2-image-captioning';
+    initialized = false;
+    constructor(imageToTextModelName, crossEncoderModelName) {
+        if (imageToTextModelName) {
+            this.imageToTextModelName = imageToTextModelName;
+        }
+        // Create the underlying cross-encoder strategy
+        this.crossEncoderReranker = new CrossEncoderRerankingStrategy(crossEncoderModelName);
+    }
+    /**
+     * Initialize the image-to-text model if not already done
+     */
+    async ensureInitialized() {
+        if (!this.initialized) {
+            try {
+                console.log(`Loading image-to-text model: ${this.imageToTextModelName}`);
+                // Set up polyfills for transformers.js
+                this.ensurePolyfills();
+                const { pipeline } = await import('@huggingface/transformers');
+                this.imageToTextModel = await pipeline('image-to-text', this.imageToTextModelName);
+                this.initialized = true;
+                console.log(`Image-to-text model loaded successfully: ${this.imageToTextModelName}`);
+            }
+            catch (error) {
+                console.warn(`Image-to-text model initialization failed: ${error instanceof Error ? error.message : 'Unknown error'}`);
+                this.isEnabled = false;
+            }
+        }
+    }
+    /**
+     * Ensure DOM polyfills are set up for transformers.js
+     */
+    ensurePolyfills() {
+        if (typeof window === 'undefined' && typeof globalThis !== 'undefined') {
+            if (typeof globalThis.self === 'undefined') {
+                globalThis.self = globalThis;
+            }
+            if (typeof global.self === 'undefined') {
+                global.self = global;
+            }
+        }
+    }
+    /**
+     * Generate text description for an image
+     */
+    async generateImageDescription(imagePath) {
+        await this.ensureInitialized();
+        if (!this.imageToTextModel) {
+            throw new Error('Image-to-text model not loaded');
+        }
+        try {
+            const result = await this.imageToTextModel(imagePath);
+            // Handle different response formats from the pipeline
+            if (Array.isArray(result) && result.length > 0) {
+                return result[0].generated_text || result[0].text || String(result[0]);
+            }
+            else if (result && typeof result === 'object') {
+                return result.generated_text || result.text || String(result);
+            }
+            else {
+                return String(result);
+            }
+        }
+        catch (error) {
+            console.warn(`Failed to generate description for image ${imagePath}: ${error instanceof Error ? error.message : 'Unknown error'}`);
+            // Fallback to filename-based description
+            const filename = imagePath.split('/').pop() || imagePath;
+            return `Image file: ${filename}`;
+        }
+    }
+    /**
+     * Rerank search results using text-derived approach
+     */
+    rerank = async (query, results, contentType) => {
+        // If strategy is disabled, return results unchanged
+        if (!this.isEnabled) {
+            return results;
+        }
+        // Validate content type
+        if (contentType && !this.supportedContentTypes.includes(contentType)) {
+            throw new Error(`Text-derived strategy does not support content type '${contentType}'. ` +
+                `Supported types: ${this.supportedContentTypes.join(', ')}`);
+        }
+        // Ensure models are initialized
+        await this.ensureInitialized();
+        // If initialization failed, return results unchanged
+        if (!this.isEnabled) {
+            console.warn('Text-derived reranker not enabled, returning results unchanged');
+            return results;
+        }
+        try {
+            // Step 1: Convert images to text descriptions
+            const processedResults = await Promise.all(results.map(async (result) => {
+                if (result.contentType === 'image') {
+                    // Generate text description for image
+                    const description = await this.generateImageDescription(result.content);
+                    return {
+                        ...result,
+                        content: description,
+                        originalContent: result.content,
+                        originalContentType: result.contentType,
+                        metadata: {
+                            ...result.metadata,
+                            originalImagePath: result.content,
+                            generatedDescription: description
+                        }
+                    };
+                }
+                return result;
+            }));
+            // Step 2: Use cross-encoder reranking on the text descriptions
+            const rerankedResults = await this.crossEncoderReranker.rerank(query, processedResults);
+            // Step 3: Restore original content for images
+            return rerankedResults.map(result => {
+                if (result.originalContent && result.originalContentType) {
+                    return {
+                        ...result,
+                        content: result.originalContent,
+                        contentType: result.originalContentType,
+                        // Keep the generated description in metadata for reference
+                        metadata: {
+                            ...result.metadata,
+                            generatedDescription: result.content
+                        }
+                    };
+                }
+                return result;
+            });
+        }
+        catch (error) {
+            console.warn(`Text-derived reranking failed: ${error instanceof Error ? error.message : 'Unknown error'}. ` +
+                `Returning original results.`);
+            return results;
+        }
+    };
+    /**
+     * Configure the reranking strategy
+     */
+    configure(config) {
+        if (config.imageToTextModel && typeof config.imageToTextModel === 'string') {
+            this.imageToTextModelName = config.imageToTextModel;
+            // Reset initialization to use new model
+            this.initialized = false;
+            this.imageToTextModel = null;
+        }
+        if (config.crossEncoderModel && typeof config.crossEncoderModel === 'string') {
+            this.crossEncoderReranker.configure({ modelName: config.crossEncoderModel });
+        }
+        if (config.enabled !== undefined) {
+            this.isEnabled = Boolean(config.enabled);
+        }
+    }
+    /**
+     * Get metadata about this reranking strategy
+     */
+    getMetadata() {
+        return {
+            description: 'Text-derived reranking that converts images to text descriptions then applies cross-encoder reranking',
+            requiredModels: [
+                'Xenova/vit-gpt2-image-captioning', // Image-to-text model
+                'Xenova/ms-marco-MiniLM-L-6-v2' // Cross-encoder model
+            ],
+            configOptions: {
+                imageToTextModel: {
+                    type: 'string',
+                    description: 'Image-to-text model name for generating descriptions',
+                    default: 'Xenova/vit-gpt2-image-captioning'
+                },
+                crossEncoderModel: {
+                    type: 'string',
+                    description: 'Cross-encoder model name for text reranking',
+                    default: 'Xenova/ms-marco-MiniLM-L-6-v2'
+                },
+                enabled: {
+                    type: 'boolean',
+                    description: 'Enable or disable text-derived reranking',
+                    default: true
+                }
+            }
+        };
+    }
+    /**
+     * Check if the strategy is ready to use
+     */
+    async isReady() {
+        await this.ensureInitialized();
+        const crossEncoderReady = await this.crossEncoderReranker.isReady();
+        return this.isEnabled && this.imageToTextModel !== null && crossEncoderReady;
+    }
+    /**
+     * Get the current model names being used
+     */
+    getModelNames() {
+        return {
+            imageToText: this.imageToTextModelName,
+            crossEncoder: this.crossEncoderReranker.getModelName()
+        };
+    }
+    /**
+     * Clean up resources
+     */
+    async cleanup() {
+        this.initialized = false;
+        this.imageToTextModel = null;
+        await this.crossEncoderReranker.cleanup();
+    }
+}
+/**
+ * Factory function to create a text-derived reranking strategy
+ */
+export function createTextDerivedStrategy(imageToTextModelName, crossEncoderModelName) {
+    return new TextDerivedRerankingStrategy(imageToTextModelName, crossEncoderModelName);
+}
+/**
+ * Create a RerankFunction using the text-derived strategy
+ */
+export function createTextDerivedRerankFunction(imageToTextModelName, crossEncoderModelName) {
+    const strategy = createTextDerivedStrategy(imageToTextModelName, crossEncoderModelName);
+    return strategy.rerank;
+}
+/**
+ * Metadata-Based Reranking Strategy Implementation
+ *
+ * Reranks search results based on filename patterns, metadata, and content type
+ * information. This strategy is particularly useful for multimodal content where
+ * semantic similarity might not capture all relevant aspects.
+ */
+export class MetadataRerankingStrategy {
+    name = 'metadata';
+    supportedContentTypes = ['text', 'image', 'pdf', 'docx'];
+    isEnabled = true;
+    config;
+    constructor(config) {
+        // Default configuration with reasonable weights and boost factors
+        const defaultConfig = {
+            weights: {
+                filename: 0.4,
+                contentType: 0.3,
+                metadata: 0.3
+            },
+            boostFactors: {
+                diagram: 1.5,
+                chart: 1.4,
+                graph: 1.4,
+                image: 1.2,
+                screenshot: 1.3,
+                figure: 1.3
+            },
+            keywordBoosts: {
+                // Technical terms
+                'api': 1.2,
+                'architecture': 1.3,
+                'design': 1.2,
+                'implementation': 1.2,
+                'configuration': 1.2,
+                'setup': 1.2,
+                'guide': 1.2,
+                'tutorial': 1.2,
+                'example': 1.1,
+                'demo': 1.1,
+                // Visual content indicators
+                'visual': 1.3,
+                'overview': 1.2,
+                'flow': 1.3,
+                'process': 1.2,
+                'workflow': 1.3
+            }
+        };
+        this.config = {
+            weights: { ...defaultConfig.weights, ...config?.weights },
+            boostFactors: { ...defaultConfig.boostFactors, ...config?.boostFactors },
+            keywordBoosts: { ...defaultConfig.keywordBoosts, ...config?.keywordBoosts }
+        };
+    }
+    /**
+     * Calculate filename-based score
+     */
+    calculateFilenameScore(query, filename) {
+        const queryLower = query.toLowerCase();
+        const filenameLower = filename.toLowerCase();
+        let score = 0;
+        // Exact filename match gets highest score
+        if (filenameLower.includes(queryLower)) {
+            score += 1.0;
+        }
+        // Word-level matching
+        const queryWords = queryLower.split(/\s+/).filter(word => word.length > 2);
+        const filenameWords = filenameLower.split(/[_\-\s\.]+/).filter(word => word.length > 2);
+        for (const queryWord of queryWords) {
+            for (const filenameWord of filenameWords) {
+                if (filenameWord.includes(queryWord) || queryWord.includes(filenameWord)) {
+                    score += 0.3;
+                }
+            }
+        }
+        // Apply keyword boosts
+        for (const [keyword, boost] of Object.entries(this.config.keywordBoosts)) {
+            if (filenameLower.includes(keyword)) {
+                score *= boost;
+            }
+        }
+        // Apply pattern-based boosts
+        for (const [pattern, boost] of Object.entries(this.config.boostFactors)) {
+            if (filenameLower.includes(pattern)) {
+                score *= boost;
+            }
+        }
+        return Math.min(score, 2.0); // Cap at 2.0 to prevent extreme scores
+    }
+    /**
+     * Calculate content type-based score
+     */
+    calculateContentTypeScore(query, contentType) {
+        const queryLower = query.toLowerCase();
+        // Base scores for different content types
+        const contentTypeScores = {
+            'image': 0.8,
+            'text': 1.0,
+            'pdf': 0.9,
+            'docx': 0.9
+        };
+        let score = contentTypeScores[contentType] || 0.5;
+        // Boost image content for visual-related queries
+        if (contentType === 'image') {
+            const visualKeywords = ['diagram', 'chart', 'graph', 'image', 'visual', 'screenshot', 'figure', 'illustration', 'visualization'];
+            for (const keyword of visualKeywords) {
+                if (queryLower.includes(keyword)) {
+                    score *= 2.0; // More aggressive boost
+                    break;
+                }
+            }
+        }
+        // Boost document content for text-heavy queries
+        if (contentType === 'text' || contentType === 'pdf' || contentType === 'docx') {
+            const textKeywords = ['documentation', 'guide', 'tutorial', 'explanation', 'description', 'details'];
+            for (const keyword of textKeywords) {
+                if (queryLower.includes(keyword)) {
+                    score *= 1.8; // More aggressive boost
+                    break;
+                }
+            }
+        }
+        return score;
+    }
+    /**
+     * Calculate metadata-based score
+     */
+    calculateMetadataScore(query, metadata = {}) {
+        let score = 0;
+        const queryLower = query.toLowerCase();
+        // Check various metadata fields
+        const metadataFields = ['title', 'description', 'tags', 'category', 'type'];
+        for (const field of metadataFields) {
+            const value = metadata[field];
+            if (typeof value === 'string') {
+                const valueLower = value.toLowerCase();
+                if (valueLower.includes(queryLower)) {
+                    score += 0.5;
+                }
+                // Word-level matching
+                const queryWords = queryLower.split(/\s+/).filter(word => word.length > 2);
+                for (const word of queryWords) {
+                    if (valueLower.includes(word)) {
+                        score += 0.2;
+                    }
+                }
+            }
+            else if (Array.isArray(value)) {
+                // Handle tag arrays
+                for (const item of value) {
+                    if (typeof item === 'string' && item.toLowerCase().includes(queryLower)) {
+                        score += 0.3;
+                    }
+                }
+            }
+        }
+        // Special handling for image metadata
+        if (metadata.dimensions) {
+            // Larger images might be more important
+            const { width, height } = metadata.dimensions;
+            if (width && height && width * height > 500000) { // > 500k pixels
+                score += 0.5; // More significant boost for large images
+            }
+        }
+        // File size considerations
+        if (metadata.fileSize) {
+            // Very small files might be less important
+            if (metadata.fileSize < 1000) {
+                score -= 0.1;
+            }
+        }
+        return Math.max(score, 0); // Ensure non-negative
+    }
+    /**
+     * Rerank search results using metadata-based scoring
+     */
+    rerank = async (query, results, contentType) => {
+        // If strategy is disabled, return results unchanged
+        if (!this.isEnabled) {
+            return results;
+        }
+        // Validate content type if specified
+        if (contentType && !this.supportedContentTypes.includes(contentType)) {
+            console.warn(`Metadata strategy does not support content type '${contentType}'. ` +
+                `Supported types: ${this.supportedContentTypes.join(', ')}. Proceeding anyway.`);
+        }
+        try {
+            // Calculate metadata scores for each result
+            const scoredResults = results.map(result => {
+                const filename = result.document.source.split('/').pop() || result.document.source;
+                // Calculate individual scores
+                const filenameScore = this.calculateFilenameScore(query, filename);
+                const contentTypeScore = this.calculateContentTypeScore(query, result.contentType);
+                const metadataScore = this.calculateMetadataScore(query, result.metadata);
+                // Combine scores using configured weights
+                const metadataBoost = (filenameScore * this.config.weights.filename +
+                    contentTypeScore * this.config.weights.contentType +
+                    metadataScore * this.config.weights.metadata);
+                // Combine with original vector similarity score
+                // Use a weighted combination where metadata boost can have significant influence
+                const combinedScore = result.score * 0.4 + metadataBoost * 0.6;
+                return {
+                    ...result,
+                    score: combinedScore,
+                    metadata: {
+                        ...result.metadata,
+                        rerankingScores: {
+                            original: result.score,
+                            filename: filenameScore,
+                            contentType: contentTypeScore,
+                            metadata: metadataScore,
+                            combined: combinedScore
+                        }
+                    }
+                };
+            });
+            // Sort by combined score (descending)
+            scoredResults.sort((a, b) => b.score - a.score);
+            return scoredResults;
+        }
+        catch (error) {
+            console.warn(`Metadata reranking failed: ${error instanceof Error ? error.message : 'Unknown error'}. ` +
+                `Returning original results.`);
+            return results;
+        }
+    };
+    /**
+     * Configure the reranking strategy
+     */
+    configure(config) {
+        if (config.weights && typeof config.weights === 'object') {
+            this.config.weights = { ...this.config.weights, ...config.weights };
+        }
+        if (config.boostFactors && typeof config.boostFactors === 'object') {
+            this.config.boostFactors = { ...this.config.boostFactors, ...config.boostFactors };
+        }
+        if (config.keywordBoosts && typeof config.keywordBoosts === 'object') {
+            this.config.keywordBoosts = { ...this.config.keywordBoosts, ...config.keywordBoosts };
+        }
+        if (config.enabled !== undefined) {
+            this.isEnabled = Boolean(config.enabled);
+        }
+    }
+    /**
+     * Get metadata about this reranking strategy
+     */
+    getMetadata() {
+        return {
+            description: 'Metadata-based reranking using filename patterns, content types, and file metadata',
+            requiredModels: [], // No models required
+            configOptions: {
+                weights: {
+                    type: 'object',
+                    description: 'Weights for different scoring components',
+                    default: this.config.weights,
+                    properties: {
+                        filename: { type: 'number', min: 0, max: 1 },
+                        contentType: { type: 'number', min: 0, max: 1 },
+                        metadata: { type: 'number', min: 0, max: 1 }
+                    }
+                },
+                boostFactors: {
+                    type: 'object',
+                    description: 'Boost factors for specific file patterns',
+                    default: this.config.boostFactors
+                },
+                keywordBoosts: {
+                    type: 'object',
+                    description: 'Boost factors for specific keywords in filenames',
+                    default: this.config.keywordBoosts
+                },
+                enabled: {
+                    type: 'boolean',
+                    description: 'Enable or disable metadata-based reranking',
+                    default: true
+                }
+            }
+        };
+    }
+    /**
+     * Check if the strategy is ready to use
+     */
+    async isReady() {
+        // Metadata strategy doesn't require model loading, so it's always ready if enabled
+        return this.isEnabled;
+    }
+    /**
+     * Get current configuration
+     */
+    getConfig() {
+        return { ...this.config };
+    }
+    /**
+     * Clean up resources (no-op for metadata strategy)
+     */
+    async cleanup() {
+        // No resources to clean up for metadata-based reranking
+    }
+}
+/**
+ * Factory function to create a metadata reranking strategy
+ */
+export function createMetadataStrategy(config) {
+    return new MetadataRerankingStrategy(config);
+}
+/**
+ * Create a RerankFunction using the metadata strategy
+ */
+export function createMetadataRerankFunction(config) {
+    const strategy = createMetadataStrategy(config);
+    return strategy.rerank;
+}
+/**
+ * Create a RerankFunction using the cross-encoder strategy
+ *
+ * This provides backward compatibility with the existing RerankFunction interface
+ * while using the new strategy-based architecture internally.
+ */
+export function createCrossEncoderRerankFunction(modelName) {
+    const strategy = createCrossEncoderStrategy(modelName);
+    return strategy.rerank;
+}
+//# sourceMappingURL=reranking-strategies.js.map