squish-memory 1.1.5 → 1.2.0

package/core/graph/llm-entity-extractor.ts

@@ -0,0 +1,402 @@
+/**
+ * LLM Entity Extractor
+ * 
+ * Extracts named entities and relationships from memory content using
+ * LLM-powered analysis. Falls back to regex extraction when LLM is unavailable.
+ */
+
+import { logger } from '../logger.js';
+import { config } from '../../config.js';
+import {
+  extractEntities as regexExtractEntities,
+  linkEntitiesToMemories,
+  type EntityType,
+  type ExtractedEntity,
+} from '../memory/entity-extractor.js';
+
+// ─── Types ──────────────────────────────────────────────────────────────────
+
+export type RelationType =
+  | 'works_on'
+  | 'depends_on'
+  | 'manages'
+  | 'uses'
+  | 'caused'
+  | 'located_in'
+  | 'belongs_to'
+  | 'reports_to'
+  | 'occurred_on'
+  | 'affects'
+  | 'contains'
+  | 'implements'
+  | 'extends'
+  | 'related_to'
+  | 'part_of'
+  | 'owns'
+  | 'created'
+  | 'resolved'
+  | 'blocks';
+
+export interface ExtractedRelation {
+  fromEntity: string;
+  toEntity: string;
+  relationType: RelationType;
+  confidence: number;
+  context: string;
+}
+
+export interface LLMExtractionResult {
+  entities: ExtractedEntity[];
+  relations: ExtractedRelation[];
+  source: 'llm' | 'regex' | 'none';
+}
+
+// ─── LLM Prompt ─────────────────────────────────────────────────────────────
+
+const ENTITY_EXTRACTION_PROMPT = `You are an entity and relationship extractor for an AI memory system. Given text, extract:
+
+1. ENTITIES: Named things mentioned (people, projects, systems, tools, concepts, events)
+2. RELATIONSHIPS: How entities connect to each other
+
+Return ONLY valid JSON in this exact format:
+{
+  "entities": [
+    {"name": "Alice", "type": "person", "confidence": 0.95},
+    {"name": "Project Atlas", "type": "concept", "confidence": 0.9}
+  ],
+  "relations": [
+    {"from": "Alice", "to": "Project Atlas", "type": "works_on", "confidence": 0.85, "context": "Alice is the tech lead on Project Atlas"},
+    {"from": "Project Atlas", "to": "PostgreSQL", "type": "uses", "confidence": 0.9, "context": "Project Atlas uses PostgreSQL for its primary datastore"}
+  ]
+}
+
+Entity types: person, file, function, class, concept, tool, date, location, pattern, technique, other
+Relation types: works_on, depends_on, manages, uses, caused, located_in, belongs_to, reports_to, occurred_on, affects, contains, implements, extends, related_to, part_of, owns, created, resolved, blocks
+
+Rules:
+- Extract only explicitly mentioned entities and relationships
+- Use the most specific relation type available
+- Include confidence scores (0-1)
+- Keep entity names as they appear in text (preserve capitalization)
+- For ambiguous relationships, use "related_to"
+- Do NOT invent entities or relationships not present in the text`;
+
+// ─── LLM Call Abstraction ───────────────────────────────────────────────────
+
+/**
+ * Call an LLM for entity extraction.
+ * Uses OpenAI-compatible API (works with OpenAI, Ollama, LM Studio, local endpoints).
+ */
+async function callLLM(prompt: string, content: string): Promise<string | null> {
+  // Determine which provider to use for extraction
+  const provider = config.llmProvider;
+  const endpoint = config.llmEndpoint || config.lmStudioUrl || config.ollamaUrl;
+  const model = config.llmExtractionModel;
+  if (!model) {
+    logger.debug('LLM extraction model not configured; using regex fallback');
+    return null;
+  }
+  
+  // Try OpenAI first if configured and provider matches
+  if ((provider === 'openai' || !provider) && config.openAiApiKey) {
+    try {
+      const chatUrl = config.openAiApiUrl.replace('/embeddings', '/chat/completions');
+      const response = await fetch(chatUrl, {
+        method: 'POST',
+        headers: {
+          'Content-Type': 'application/json',
+          Authorization: `Bearer ${config.openAiApiKey}`,
+        },
+        body: JSON.stringify({
+          model,
+          messages: [
+            { role: 'system', content: prompt },
+            { role: 'user', content },
+          ],
+          temperature: 0.1,
+          max_tokens: 2000,
+        }),
+      });
+
+      if (!response.ok) {
+        logger.warn(`LLM entity extraction failed (OpenAI): ${response.status}`);
+        return null;
+      }
+
+      const payload = await response.json() as { choices?: Array<{ message?: { content?: string } }> };
+      return payload.choices?.[0]?.message?.content ?? null;
+    } catch (error) {
+      logger.warn('LLM entity extraction error (OpenAI):', { error: error as Error });
+    }
+  }
+
+  // Try Ollama if configured and provider matches
+  if ((provider === 'ollama' || !provider) && config.ollamaUrl) {
+    try {
+      const response = await fetch(`${config.ollamaUrl}/api/chat`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({
+          model,
+          messages: [
+            { role: 'system', content: prompt },
+            { role: 'user', content },
+          ],
+          stream: false,
+          options: { temperature: 0.1 },
+        }),
+      });
+
+      if (!response.ok) {
+        logger.warn(`Ollama entity extraction failed: ${response.status}`);
+        return null;
+      }
+
+      const payload = await response.json() as { message?: { content?: string } };
+      return payload.message?.content ?? null;
+    } catch (error) {
+      logger.warn('LLM entity extraction error (Ollama):', { error: error as Error });
+    }
+  }
+
+  // Try LM Studio / custom endpoint if configured and provider matches
+  if ((provider === 'lmstudio' || provider === 'local' || !provider) && (config.lmStudioUrl || endpoint)) {
+    try {
+      const baseUrl = endpoint || config.lmStudioUrl;
+      if (!baseUrl) return null;
+      const response = await fetch(`${baseUrl}/v1/chat/completions`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({
+          model,
+          messages: [
+            { role: 'system', content: prompt },
+            { role: 'user', content },
+          ],
+          temperature: 0.1,
+          max_tokens: 2000,
+        }),
+      });
+
+      if (!response.ok) {
+        logger.warn(`LM Studio entity extraction failed: ${response.status}`);
+        return null;
+      }
+
+      const payload = await response.json() as { choices?: Array<{ message?: { content?: string } }> };
+      return payload.choices?.[0]?.message?.content ?? null;
+    } catch (error) {
+      logger.warn('LLM entity extraction error (LM Studio):', { error: error as Error });
+    }
+  }
+
+  return null;
+}
+
+// ─── JSON Parsing ────────────────────────────────────────────────────────────
+
+/**
+ * Parse LLM response into structured extraction result.
+ * Handles various response formats (raw JSON, markdown code blocks, etc.)
+ */
+function parseLLMResponse(response: string): { entities: ExtractedEntity[]; relations: ExtractedRelation[] } | null {
+  // Try to extract JSON from the response
+  let jsonStr = response.trim();
+
+  // Handle markdown code blocks
+  const codeBlockMatch = jsonStr.match(/```(?:json)?\s*([\s\S]*?)```/);
+  if (codeBlockMatch) {
+    jsonStr = codeBlockMatch[1].trim();
+  }
+
+  // Try to find JSON object in the response
+  const jsonMatch = jsonStr.match(/\{[\s\S]*\}/);
+  if (!jsonMatch) {
+    logger.debug('No JSON found in LLM response');
+    return null;
+  }
+
+  try {
+    const parsed = JSON.parse(jsonMatch[0]);
+
+    // Validate structure
+    if (!parsed.entities || !Array.isArray(parsed.entities)) {
+      logger.debug('LLM response missing entities array');
+      return null;
+    }
+
+    const entities: ExtractedEntity[] = (parsed.entities as any[])
+      .filter((e: any) => e.name && e.type)
+      .map((e: any) => ({
+        name: String(e.name),
+        type: validateEntityType(String(e.type)),
+        confidence: typeof e.confidence === 'number' ? Math.min(1, Math.max(0, e.confidence)) : 0.8,
+        startIndex: 0,
+        endIndex: 0,
+        context: String(e.context || ''),
+        normalized: e.normalized ? String(e.normalized) : undefined,
+      }));
+
+    const relations: ExtractedRelation[] = (parsed.relations || [])
+      .filter((r: any) => r.from && r.to && r.type)
+      .map((r: any) => ({
+        fromEntity: String(r.from),
+        toEntity: String(r.to),
+        relationType: validateRelationType(String(r.type)),
+        confidence: typeof r.confidence === 'number' ? Math.min(1, Math.max(0, r.confidence)) : 0.7,
+        context: String(r.context || ''),
+      }));
+
+    return { entities, relations };
+  } catch (error) {
+    logger.debug('Failed to parse LLM response as JSON', { error: error as Error });
+    return null;
+  }
+}
+
+function validateEntityType(type: string): EntityType {
+  const validTypes: EntityType[] = ['person', 'file', 'function', 'class', 'concept', 'tool', 'date', 'location', 'pattern', 'technique', 'other'];
+  const lower = type.toLowerCase() as EntityType;
+  if (validTypes.includes(lower)) return lower;
+  return 'other';
+}
+
+function validateRelationType(type: string): RelationType {
+  const validTypes: RelationType[] = [
+    'works_on', 'depends_on', 'manages', 'uses', 'caused', 'located_in',
+    'belongs_to', 'reports_to', 'occurred_on', 'affects', 'contains',
+    'implements', 'extends', 'related_to', 'part_of', 'owns', 'created',
+    'resolved', 'blocks',
+  ];
+  const lower = type.toLowerCase() as RelationType;
+  if (validTypes.includes(lower)) return lower;
+  return 'related_to';
+}
+
+// ─── Main Extraction Functions ───────────────────────────────────────────────
+
+/**
+ * Extract entities and relationships from text using LLM.
+ * Falls back to regex extraction when LLM is unavailable.
+ */
+export async function extractEntitiesAndRelations(
+  content: string,
+  options?: {
+    preferLLM?: boolean;
+    maxContentLength?: number;
+  }
+): Promise<LLMExtractionResult> {
+  const { preferLLM = true, maxContentLength = 4000 } = options || {};
+
+  // Truncate very long content
+  const truncatedContent = content.length > maxContentLength
+    ? content.substring(0, maxContentLength) + '...'
+    : content;
+
+  // Try LLM extraction first if preferred
+  if (preferLLM) {
+    const llmResult = await extractWithLLM(truncatedContent);
+    if (llmResult) {
+      logger.debug('LLM entity extraction succeeded', {
+        entityCount: llmResult.entities.length,
+        relationCount: llmResult.relations.length,
+      });
+      return { ...llmResult, source: 'llm' };
+    }
+  }
+
+  // Fall back to regex extraction
+  const regexEntities = await regexExtractEntities(content);
+  if (regexEntities.length > 0) {
+    logger.debug('Regex entity extraction fallback', {
+      entityCount: regexEntities.length,
+    });
+    return {
+      entities: regexEntities,
+      relations: [], // Regex extraction doesn't produce relations
+      source: 'regex',
+    };
+  }
+
+  return { entities: [], relations: [], source: 'none' };
+}
+
+/**
+ * Extract entities and relations using LLM.
+ */
+async function extractWithLLM(content: string): Promise<Omit<LLMExtractionResult, 'source'> | null> {
+  const response = await callLLM(ENTITY_EXTRACTION_PROMPT, content);
+  if (!response) return null;
+
+  const parsed = parseLLMResponse(response);
+  if (!parsed) return null;
+
+  // Validate that extracted entities actually appear in the content
+  const validatedEntities = parsed.entities.filter((entity) => {
+    // Allow entities whose name or a significant substring appears in content
+    const nameLower = entity.name.toLowerCase();
+    const contentLower = content.toLowerCase();
+    // Check if the entity name or a key part of it appears in content
+    const words = nameLower.split(/\s+/);
+    return words.some(word => word.length > 2 && contentLower.includes(word));
+  });
+
+  // Validate that relation entities exist in the entity list
+  const entityNames = new Set(validatedEntities.map(e => e.name.toLowerCase()));
+  const validatedRelations = parsed.relations.filter((relation) => {
+    const fromExists = entityNames.has(relation.fromEntity.toLowerCase());
+    const toExists = entityNames.has(relation.toEntity.toLowerCase());
+    // Allow relations where at least one entity is known
+    // (the LLM might use slightly different names)
+    return fromExists || toExists || 
+      // Also allow if both entity names appear in the content
+      (content.toLowerCase().includes(relation.fromEntity.toLowerCase()) &&
+       content.toLowerCase().includes(relation.toEntity.toLowerCase()));
+  });
+
+  return {
+    entities: validatedEntities,
+    relations: validatedRelations,
+  };
+}
+
+/**
+ * Extract entities and relations from multiple memories in batch.
+ * More efficient than calling extractEntitiesAndRelations for each memory.
+ */
+export async function batchExtractEntitiesAndRelations(
+  contents: string[],
+  options?: {
+    preferLLM?: boolean;
+    maxContentLength?: number;
+    batchSize?: number;
+  }
+): Promise<LLMExtractionResult[]> {
+  const { batchSize = 5 } = options || {};
+  const results: LLMExtractionResult[] = [];
+
+  // Process in batches to avoid overwhelming the LLM
+  for (let i = 0; i < contents.length; i += batchSize) {
+    const batch = contents.slice(i, i + batchSize);
+    const batchResults = await Promise.all(
+      batch.map(content => extractEntitiesAndRelations(content, options))
+    );
+    results.push(...batchResults);
+  }
+
+  return results;
+}
+
+/**
+ * Get the extraction prompt for testing/debugging purposes.
+ */
+export function getExtractionPrompt(): string {
+  return ENTITY_EXTRACTION_PROMPT;
+}
+
+/**
+ * Parse an LLM response for testing/debugging purposes.
+ */
+export function testParseLLMResponse(response: string): { entities: ExtractedEntity[]; relations: ExtractedRelation[] } | null {
+  return parseLLMResponse(response);
+}