@4da/mcp-server 5.0.0 → 5.0.2

package/dist/llm.js

@@ -1,502 +0,0 @@
-// SPDX-License-Identifier: Apache-2.0
-/**
- * LLM Client for 4DA MCP Server
- *
- * Provides intelligence synthesis over raw tool data.
- * Supports Anthropic Claude, OpenAI, and Ollama (local).
- *
- * Features:
- * - Tiered model selection (use smaller models for simpler tasks)
- * - Optimized prompts for local models
- * - Automatic fallback
- */
-// =============================================================================
-// Model Tiers - Optimized for 16GB VRAM (RTX 4080 Super)
-// =============================================================================
-/**
- * Default models for different task complexities
- * Optimized for 16GB VRAM
- */
-export const DEFAULT_OLLAMA_MODELS = {
-    light: "gemma3:4b",
-    medium: "qwen3:8b",
-    heavy: "qwen3:14b",
-};
-/**
- * Task complexity by tool
- */
-export const TOOL_COMPLEXITY = {
-    daily_briefing: "medium", // Needs to synthesize multiple themes
-    score_autopsy: "heavy", // Needs to reason about score correctness
-    trend_analysis: "heavy", // Needs pattern recognition + prediction
-    context_analysis: "medium", // Structured advice
-    topic_connections: "heavy", // Graph reasoning
-};
-// =============================================================================
-// Configuration
-// =============================================================================
-import { readFileSync, existsSync, statSync } from "fs";
-import { join, dirname } from "path";
-import { fileURLToPath } from "node:url";
-const __filename = fileURLToPath(import.meta.url);
-const __dirname = dirname(__filename);
-/**
- * Find settings.json by looking in common locations
- */
-function findSettingsFile() {
-    // 1. Environment variable (highest priority)
-    const envPath = process.env.FOURDA_SETTINGS_PATH;
-    if (envPath && existsSync(envPath)) {
-        return envPath;
-    }
-    // 2. Sibling to database path
-    const dbPath = process.env.FOURDA_DB_PATH;
-    if (dbPath) {
-        const settingsPath = join(dirname(dbPath), "settings.json");
-        if (existsSync(settingsPath)) {
-            return settingsPath;
-        }
-    }
-    // 3. Common relative locations (portable — no hardcoded dev paths)
-    const searchPaths = [
-        join(process.cwd(), "data", "settings.json"), // CWD/data/
-        join(__dirname, "..", "..", "data", "settings.json"), // Project root (mcp-4da-server/../data/)
-        join(process.cwd(), "..", "data", "settings.json"), // Parent/data/
-    ];
-    for (const candidate of searchPaths) {
-        if (existsSync(candidate)) {
-            return candidate;
-        }
-    }
-    return null;
-}
-// =============================================================================
-// Settings Cache — re-read only when file mtime changes
-// =============================================================================
-let _cachedConfig = null;
-let _cachedSettingsPath = null;
-let _cachedMtimeMs = 0;
-/**
- * Get LLM configuration from settings.json file.
- *
- * Uses a cached approach: the file is only re-read if its mtime has changed,
- * avoiding unnecessary filesystem reads on every tool invocation.
- */
-export function getLLMConfig(_db) {
-    const settingsPath = findSettingsFile();
-    if (!settingsPath) {
-        // Return environment-based config as fallback
-        return {
-            provider: process.env.LLM_PROVIDER || null,
-            anthropic_api_key: process.env.ANTHROPIC_API_KEY,
-            openai_api_key: process.env.OPENAI_API_KEY,
-            ollama_url: process.env.OLLAMA_URL || "http://localhost:11434",
-            model: process.env.LLM_MODEL,
-            model_light: process.env.LLM_MODEL_LIGHT,
-            model_heavy: process.env.LLM_MODEL_HEAVY,
-        };
-    }
-    // Check if we can use the cached config
-    try {
-        const currentMtimeMs = statSync(settingsPath).mtimeMs;
-        if (_cachedConfig &&
-            _cachedSettingsPath === settingsPath &&
-            _cachedMtimeMs === currentMtimeMs) {
-            return _cachedConfig;
-        }
-        const settings = JSON.parse(readFileSync(settingsPath, "utf-8"));
-        const llm = settings.llm || {};
-        const config = {
-            provider: (llm.provider ?? null),
-            anthropic_api_key: llm.api_key || llm.anthropic_api_key,
-            openai_api_key: llm.openai_api_key,
-            ollama_url: llm.base_url || llm.ollama_url || "http://localhost:11434",
-            model: llm.model,
-            model_light: llm.model_light,
-            model_heavy: llm.model_heavy,
-        };
-        // Update cache
-        _cachedConfig = config;
-        _cachedSettingsPath = settingsPath;
-        _cachedMtimeMs = currentMtimeMs;
-        return config;
-    }
-    catch (error) {
-        console.error("Failed to read settings.json:", error);
-        return { provider: null };
-    }
-}
-/**
- * Check if LLM synthesis is available
- */
-export function canSynthesize(config) {
-    if (!config.provider)
-        return false;
-    switch (config.provider) {
-        case "anthropic":
-            return !!config.anthropic_api_key;
-        case "openai":
-            return !!config.openai_api_key;
-        case "ollama":
-            return true; // Ollama doesn't need a key
-        default:
-            return false;
-    }
-}
-/**
- * Get the appropriate model for task complexity
- */
-function getModelForComplexity(config, complexity) {
-    if (config.provider !== "ollama") {
-        // Cloud providers use the configured model
-        return config.model || (config.provider === "anthropic" ? "claude-sonnet-4-6" : "gpt-4o");
-    }
-    // Ollama: use tiered models
-    switch (complexity) {
-        case "light":
-            return config.model_light || config.model || DEFAULT_OLLAMA_MODELS.light;
-        case "medium":
-            return config.model || DEFAULT_OLLAMA_MODELS.medium;
-        case "heavy":
-            return config.model_heavy || config.model || DEFAULT_OLLAMA_MODELS.heavy;
-    }
-}
-// =============================================================================
-// Synthesis
-// =============================================================================
-/**
- * Synthesize insights using LLM
- *
- * Accepts an optional AbortSignal to allow timeout/cancellation of the
- * underlying fetch calls.
- */
-export async function synthesize(config, request, signal) {
-    if (!canSynthesize(config)) {
-        throw new Error("LLM not configured - cannot synthesize");
-    }
-    const startTime = Date.now();
-    const complexity = request.complexity || "medium";
-    let result;
-    switch (config.provider) {
-        case "anthropic":
-            result = await synthesizeAnthropic(config, request, signal);
-            break;
-        case "openai":
-            result = await synthesizeOpenAI(config, request, signal);
-            break;
-        case "ollama":
-            result = await synthesizeOllama(config, request, complexity, signal);
-            break;
-        default:
-            throw new Error(`Unknown LLM provider: ${config.provider}`);
-    }
-    result.latency_ms = Date.now() - startTime;
-    return result;
-}
-/**
- * Synthesize with a timeout wrapper.
- *
- * Wraps the LLM API call with an AbortController + configurable timeout.
- * On timeout, throws a descriptive error so callers can fall back to
- * data-only results.
- *
- * @param config  - LLM provider configuration
- * @param request - Synthesis request (system prompt, user prompt, etc.)
- * @param timeoutMs - Timeout in milliseconds (default: 30000)
- */
-export async function synthesizeWithTimeout(config, request, timeoutMs = 30000) {
-    const controller = new AbortController();
-    const timeout = setTimeout(() => controller.abort(), timeoutMs);
-    try {
-        return await synthesize(config, request, controller.signal);
-    }
-    catch (error) {
-        if (error instanceof Error && error.name === "AbortError") {
-            throw new Error(`LLM timeout after ${timeoutMs / 1000}s`);
-        }
-        throw error;
-    }
-    finally {
-        clearTimeout(timeout);
-    }
-}
-/**
- * Anthropic Claude synthesis
- */
-async function synthesizeAnthropic(config, request, signal) {
-    const model = config.model || "claude-sonnet-4-6";
-    const maxTokens = request.max_tokens || 1024;
-    const response = await fetch("https://api.anthropic.com/v1/messages", {
-        method: "POST",
-        headers: {
-            "Content-Type": "application/json",
-            "x-api-key": config.anthropic_api_key,
-            "anthropic-version": "2023-06-01",
-        },
-        body: JSON.stringify({
-            model,
-            max_tokens: maxTokens,
-            system: request.system,
-            messages: [{ role: "user", content: request.prompt }],
-        }),
-        signal,
-    });
-    if (!response.ok) {
-        const error = await response.text();
-        throw new Error(`Anthropic API error: ${error}`);
-    }
-    const data = (await response.json());
-    return {
-        synthesis: data.content[0]?.text || "",
-        model_used: model,
-        tokens_used: data.usage?.output_tokens,
-    };
-}
-/**
- * OpenAI synthesis
- */
-async function synthesizeOpenAI(config, request, signal) {
-    const model = config.model || "gpt-4o";
-    const maxTokens = request.max_tokens || 1024;
-    const response = await fetch("https://api.openai.com/v1/chat/completions", {
-        method: "POST",
-        headers: {
-            "Content-Type": "application/json",
-            Authorization: `Bearer ${config.openai_api_key}`,
-        },
-        body: JSON.stringify({
-            model,
-            max_tokens: maxTokens,
-            messages: [
-                { role: "system", content: request.system },
-                { role: "user", content: request.prompt },
-            ],
-        }),
-        signal,
-    });
-    if (!response.ok) {
-        const error = await response.text();
-        throw new Error(`OpenAI API error: ${error}`);
-    }
-    const data = (await response.json());
-    return {
-        synthesis: data.choices[0]?.message?.content || "",
-        model_used: model,
-        tokens_used: data.usage?.completion_tokens,
-    };
-}
-/**
- * Ollama synthesis (local) - optimized for local models
- */
-async function synthesizeOllama(config, request, complexity, signal) {
-    const model = getModelForComplexity(config, complexity);
-    const baseUrl = config.ollama_url || "http://localhost:11434";
-    // Use chat endpoint for better instruction following
-    const response = await fetch(`${baseUrl}/api/chat`, {
-        method: "POST",
-        headers: {
-            "Content-Type": "application/json",
-        },
-        body: JSON.stringify({
-            model,
-            messages: [
-                { role: "system", content: request.system },
-                { role: "user", content: request.prompt },
-            ],
-            stream: false,
-            options: {
-                num_predict: request.max_tokens || 512,
-                temperature: 0.7,
-                top_p: 0.9,
-            },
-        }),
-        signal,
-    });
-    if (!response.ok) {
-        const error = await response.text();
-        throw new Error(`Ollama API error: ${error}`);
-    }
-    const data = (await response.json());
-    return {
-        synthesis: data.message?.content || "",
-        model_used: model,
-        tokens_used: data.eval_count,
-    };
-}
-// =============================================================================
-// Synthesis Prompts - Optimized for Local Models
-// =============================================================================
-/**
- * Prompts are optimized for:
- * 1. Clear structure (local models need explicit formatting)
- * 2. Concise output (fewer tokens = faster inference)
- * 3. Specific instructions (reduce ambiguity)
- */
-export const SYNTHESIS_PROMPTS = {
-    /**
-     * Daily Briefing - Executive synthesis
-     * Complexity: MEDIUM
-     */
-    dailyBriefing: {
-        complexity: "medium",
-        system: `You are an intelligence analyst. Synthesize data into actionable briefings.
-
-OUTPUT FORMAT:
-1. One-sentence summary of the day
-2. The single most important thing to review (with reason)
-3. One recommended action
-
-Be direct. No filler. Every word must add value.`,
-        buildPrompt: (data, context) => `
-DATA:
-${JSON.stringify(data, null, 2)}
-
-USER CONTEXT:
-${JSON.stringify(context, null, 2)}
-
-Generate briefing with:
-1. ONE-LINE SUMMARY
-2. TOP PRIORITY (title + why it matters to this user)
-3. RECOMMENDED ACTION`,
-    },
-    /**
-     * Score Autopsy - Explain the reasoning
-     * Complexity: HEAVY (needs to reason about correctness)
-     */
-    scoreAutopsy: {
-        complexity: "heavy",
-        system: `You are a scoring forensics expert. Analyze if relevance scores are correct.
-
-OUTPUT FORMAT:
-1. VERDICT: [ACCURATE / TOO HIGH / TOO LOW]
-2. REASONING: Why you reached this verdict (cite specific data)
-3. KEY FACTOR: The main thing driving this score
-4. SUGGESTION: How to get more/fewer items like this
-
-Be analytical. Cite numbers from the data.`,
-        buildPrompt: (data, context) => `
-SCORE DATA:
-${JSON.stringify(data, null, 2)}
-
-USER INTERESTS:
-${JSON.stringify(context, null, 2)}
-
-Analyze: Is this score correct given the user's interests?`,
-    },
-    /**
-     * Trend Analysis - Predict and explain
-     * Complexity: HEAVY (needs pattern recognition + prediction)
-     */
-    trendAnalysis: {
-        complexity: "heavy",
-        system: `You are a tech trend analyst. Interpret patterns and predict trajectories.
-
-OUTPUT FORMAT:
-1. KEY SIGNAL: The most important trend (one sentence)
-2. WHY IT'S HAPPENING: Brief explanation
-3. PREDICTION: What to expect in the next 1-2 weeks
-4. ACTION: Should the user adjust their focus?
-
-Ground predictions in the data. Don't speculate wildly.`,
-        buildPrompt: (data, context) => `
-TREND DATA:
-${JSON.stringify(data, null, 2)}
-
-USER CONTEXT:
-${JSON.stringify(context, null, 2)}
-
-Interpret: What do these trends mean for this specific user?`,
-    },
-    /**
-     * Context Analysis - Personalized optimization advice
-     * Complexity: MEDIUM (structured recommendations)
-     */
-    contextAnalysis: {
-        complexity: "medium",
-        system: `You are a personalization expert. Help users optimize their content filtering.
-
-OUTPUT FORMAT:
-1. BIGGEST GAP: What's missing from their context?
-2. HIGHEST IMPACT CHANGE: One specific thing to add/remove
-3. QUICK WIN: Something they can do in 30 seconds
-
-Be specific. Give exact examples they can copy-paste.`,
-        buildPrompt: (data) => `
-CONTEXT DATA:
-${JSON.stringify(data, null, 2)}
-
-Provide optimization recommendations. Be specific and actionable.`,
-    },
-    /**
-     * Topic Connections - Explain the knowledge graph
-     * Complexity: HEAVY (needs graph reasoning)
-     */
-    topicConnections: {
-        complexity: "heavy",
-        system: `You are a knowledge graph analyst. Find hidden patterns in topic relationships.
-
-OUTPUT FORMAT:
-1. SURPRISING CONNECTION: A non-obvious relationship and its significance
-2. MISSING LINK: Topics that should connect but don't
-3. EXPLORATION PATH: A topic worth investigating based on the graph
-4. PATTERN: What the graph structure reveals about their interests
-
-Look for insights the user wouldn't notice themselves.`,
-        buildPrompt: (data, context) => `
-GRAPH DATA:
-${JSON.stringify(data, null, 2)}
-
-USER CONTEXT:
-${JSON.stringify(context, null, 2)}
-
-Interpret: What hidden patterns does this graph reveal?`,
-    },
-};
-// =============================================================================
-// Utility: Check Ollama availability
-// =============================================================================
-/**
- * Check if Ollama is running and has required models
- */
-export async function checkOllamaStatus(baseUrl = "http://localhost:11434") {
-    try {
-        const response = await fetch(`${baseUrl}/api/tags`, {
-            method: "GET",
-            signal: AbortSignal.timeout(5000),
-        });
-        if (!response.ok) {
-            return { available: false, models: [], error: "Ollama not responding" };
-        }
-        const data = (await response.json());
-        const models = data.models?.map(m => m.name) || [];
-        return { available: true, models };
-    }
-    catch (error) {
-        return {
-            available: false,
-            models: [],
-            error: error instanceof Error ? error.message : "Connection failed",
-        };
-    }
-}
-/**
- * Recommended models for 4DA synthesis tasks
- */
-export const RECOMMENDED_MODELS = {
-    "16GB_VRAM": {
-        primary: "qwen3:14b",
-        fallback: "gemma3:12b",
-        install: "ollama pull qwen3:14b",
-    },
-    "12GB_VRAM": {
-        primary: "gemma3:12b",
-        fallback: "qwen3:8b",
-        install: "ollama pull gemma3:12b",
-    },
-    "8GB_VRAM": {
-        primary: "qwen3:8b",
-        fallback: "gemma3:4b",
-        install: "ollama pull qwen3:8b",
-    },
-};
-//# sourceMappingURL=llm.js.map

package/dist/output-manager.d.ts

@@ -1,156 +0,0 @@
-/**
- * Output Manager for Dynamic Context Discovery
- *
- * Reduces token usage by ~80% through:
- * - Writing full results to files (referenced by path)
- * - Returning compact summaries inline
- * - Providing retrieval hints for common queries
- *
- * Full results stored at: ~/.local/share/4da/context/results/
- */
-/**
- * Compact result format - what gets returned inline to Claude
- */
-export interface CompactResult<T> {
-    /** 2-3 sentence human-readable summary */
-    summary: string;
-    /** Essential 20% of data inline for immediate use */
-    key_data: T;
-    /** Path to complete JSON file */
-    full_result_path: string;
-    /** Hints for retrieving specific data (jq patterns, grep patterns) */
-    retrieval_hints: string[];
-}
-/**
- * Get the results directory path (for documentation)
- */
-export declare function getResultsDir(): string;
-/**
- * Daily Briefing compact output
- */
-export interface DailyBriefingKeyData {
-    total_items: number;
-    high_priority: number;
-    top_theme: string;
-    tldr: string;
-    executive_brief?: string;
-}
-export declare function createDailyBriefingCompact(result: {
-    summary: {
-        total_items_analyzed: number;
-        high_relevance_count: number;
-        tldr: string;
-    };
-    themes: {
-        name: string;
-        item_count: number;
-    }[];
-    ai_synthesis?: {
-        executive_brief: string;
-    };
-}): CompactResult<DailyBriefingKeyData>;
-/**
- * Trend Analysis compact output
- */
-export interface TrendAnalysisKeyData {
-    total_items: number;
-    volume_trend: string;
-    top_rising: string | null;
-    top_falling: string | null;
-    anomaly_count: number;
-    narrative?: string;
-}
-export declare function createTrendAnalysisCompact(result: {
-    summary: {
-        total_items: number;
-        volume_trend: string;
-        top_rising_topic: string | null;
-        top_falling_topic: string | null;
-    };
-    anomalies: unknown[];
-    ai_interpretation?: {
-        narrative: string;
-    };
-}): CompactResult<TrendAnalysisKeyData>;
-/**
- * Score Autopsy compact output
- */
-export interface ScoreAutopsyKeyData {
-    item_title: string;
-    final_score: number;
-    score_assessment?: "accurate" | "too_high" | "too_low" | "uncertain";
-    top_contributor: string;
-    verdict?: string;
-}
-export declare function createScoreAutopsyCompact(result: {
-    item: {
-        title: string;
-    };
-    final_score: number;
-    components: {
-        name: string;
-        contribution: number;
-    }[];
-    ai_analysis?: {
-        score_assessment: "accurate" | "too_high" | "too_low" | "uncertain";
-        verdict: string;
-    };
-}): CompactResult<ScoreAutopsyKeyData>;
-/**
- * Context Analysis compact output
- */
-export interface ContextAnalysisKeyData {
-    quality_score: number;
-    interests_count: number;
-    tech_count: number;
-    gap_count: number;
-    top_recommendation: string;
-    personalized_advice?: string;
-}
-export declare function createContextAnalysisCompact(result: {
-    context_summary: {
-        interests_count: number;
-        tech_stack_count: number;
-    };
-    quality_score: {
-        overall: number;
-    };
-    coverage_gaps: unknown[];
-    recommendations: {
-        action: string;
-    }[];
-    ai_advice?: {
-        personalized_recommendations: string;
-    };
-}): CompactResult<ContextAnalysisKeyData>;
-/**
- * Topic Connections compact output
- */
-export interface TopicConnectionsKeyData {
-    items_analyzed: number;
-    node_count: number;
-    edge_count: number;
-    central_topics: string[];
-    top_connection: {
-        from: string;
-        to: string;
-        strength: number;
-    } | null;
-    graph_narrative?: string;
-}
-export declare function createTopicConnectionsCompact(result: {
-    analysis_period: {
-        items_analyzed: number;
-    };
-    nodes: unknown[];
-    edges: {
-        from: string;
-        to: string;
-        co_occurrences: number;
-    }[];
-    central_topics: string[];
-    ai_insights?: {
-        graph_narrative: string;
-    };
-}): CompactResult<TopicConnectionsKeyData>;
-//# sourceMappingURL=output-manager.d.ts.map