memorix 1.1.7 → 1.1.8

package/src/llm/quality.ts

@@ -1,248 +1,248 @@
-/**
- * LLM Quality Enhancements
- *
- * Premium memory quality features powered by LLM:
- * 1. Narrative Compression — compress verbose narratives into concise core knowledge
- * 2. Search Reranking — rerank search results by relevance to current task context
- *
- * Both features gracefully degrade: when LLM is not configured, they return
- * the original data unchanged.
- *
- * Performance targets:
- * - Compression: ~60% token reduction per stored memory
- * - Reranking: ~40% improvement in Top-5 precision
- */
-
-import { callLLM, isLLMEnabled } from './provider.js';
-
-// ── Narrative Compression ────────────────────────────────────────
-
-const COMPRESS_PROMPT = `You are a memory compression engine for a coding assistant.
-
-Compress the given narrative while preserving ALL technical facts and reasoning.
-
-Rules:
-- Remove: filler words, debugging journey, repeated info already in facts
-- Keep: specific values, file paths, error messages, version numbers, config keys, causal relationships, design reasoning
-- Merge related points into dense sentences
-- If facts are provided separately, do NOT repeat them in the compressed narrative
-- Output the compressed text ONLY, no explanation or wrapper
-
-Examples:
-Input: "我在调试过程中发现JWT token的refresh机制存在问题，具体来说是因为服务端没有实现自动续签，导致用户在24小时后会遇到静默的认证失败，之前我一直以为是网络问题但后来排查发现是token过期了"
-Output: "JWT refresh无自动续签→24h后静默认证失败（非网络问题）"
-
-Input: "Final deployment model for shadcn-blog is stable: GitHub Actions build locally, SCP artifacts to VPS, systemd manages the process. Docker was considered but rejected due to complexity overhead for a simple blog. The whole pipeline takes about 2 minutes from push to live."
-Output: "shadcn-blog部署: GH Actions构建→SCP到VPS→systemd管理, 弃Docker(复杂度过高), push到上线~2min"`;
-
-/** Gentler prompt for high-value types where reasoning context matters */
-const COMPRESS_PROMPT_GENTLE = `You are a memory compression engine for a coding assistant.
-
-Lightly compress the given narrative — preserve reasoning, trade-offs, and "why" context.
-
-Rules:
-- Only remove: obvious filler, debugging detours, info already in the separate facts list
-- PRESERVE: design reasoning, rejected alternatives, trade-off analysis, causal chains
-- Aim for ~70-80% of original length, NOT aggressive compression
-- If facts are provided separately, do NOT repeat them in the compressed narrative
-- Output the compressed text ONLY, no explanation or wrapper`;
-
-/**
- * Compress a narrative to its essential core using LLM.
- *
- * Returns the original narrative if:
- * - LLM is not enabled
- * - Narrative is already short (≤80 chars)
- * - Narrative is already concise (commands, file paths, git operations)
- * - LLM call fails
- */
-export async function compressNarrative(
-  narrative: string,
-  facts?: string[],
-  type?: string,
-): Promise<{ compressed: string; saved: number; usedLLM: boolean }> {
-  const originalTokens = estimateTokens(narrative);
-
-  // Skip compression for short narratives (≤150 chars is already concise)
-  if (!isLLMEnabled() || narrative.length <= 150) {
-    return { compressed: narrative, saved: 0, usedLLM: false };
-  }
-
-  // Skip compression for already-concise content that LLM can't meaningfully compress
-  if (shouldSkipCompression(narrative, type)) {
-    return { compressed: narrative, saved: 0, usedLLM: false };
-  }
-
-  try {
-    const factsContext = facts && facts.length > 0
-      ? `\n\nSeparate facts (already stored, don't repeat): ${facts.join('; ')}`
-      : '';
-
-    // Use gentler compression for high-value types where reasoning matters
-    const HIGH_VALUE_TYPES = new Set(['decision', 'trade-off', 'why-it-exists', 'how-it-works']);
-    const prompt = (type && HIGH_VALUE_TYPES.has(type)) ? COMPRESS_PROMPT_GENTLE : COMPRESS_PROMPT;
-    const response = await callLLM(prompt, narrative + factsContext);
-    const compressed = response.content.trim();
-
-    // Sanity check: compressed should be shorter and non-empty
-    if (!compressed || compressed.length >= narrative.length) {
-      return { compressed: narrative, saved: 0, usedLLM: true };
-    }
-
-    const compressedTokens = estimateTokens(compressed);
-    return {
-      compressed,
-      saved: originalTokens - compressedTokens,
-      usedLLM: true,
-    };
-  } catch {
-    return { compressed: narrative, saved: 0, usedLLM: false };
-  }
-}
-
-// ── Search Reranking ─────────────────────────────────────────────
-
-/** Minimal search result for reranking */
-export interface RerankCandidate {
-  id: string;
-  title: string;
-  type: string;
-  score: number;
-  narrative?: string;
-}
-
-const RERANK_PROMPT = `You are a memory relevance ranker for a coding assistant.
-
-Given a QUERY (what the user/agent is looking for) and a list of CANDIDATE memories,
-rerank them by relevance to the query.
-
-Rules:
-- Consider semantic relevance, not just keyword overlap
-- Gotchas and decisions related to the query topic should rank higher
-- Recent problem-solutions for the same component should rank higher
-- Command or audit-log memories (titles starting with "Ran:" or "Command:") should rank lower for natural-language questions unless the query is explicitly about commands, scripts, or audit history
-- Generic or loosely related memories should rank lower
-- Output ONLY a JSON array of IDs in order of relevance (most relevant first)
-- Include ALL candidate IDs, just reorder them
-
-Example output: ["r1", "r3", "r2"]`;
-
-/**
- * Rerank search results using LLM contextual understanding.
- *
- * Takes Orama's initial ranking and improves it by considering
- * semantic relevance to the current query/task context.
- *
- * Returns original order if LLM is not enabled or call fails.
- */
-export async function rerankResults(
-  query: string,
-  candidates: RerankCandidate[],
-): Promise<{ reranked: RerankCandidate[]; usedLLM: boolean }> {
-  // Skip if too few results or LLM not available
-  if (!isLLMEnabled() || candidates.length <= 2) {
-    return { reranked: candidates, usedLLM: false };
-  }
-
-  // Only rerank top-N to save LLM tokens (reranking 20+ is wasteful)
-  const MAX_RERANK = 10;
-  const toRerank = candidates.slice(0, MAX_RERANK);
-  const rest = candidates.slice(MAX_RERANK);
-
-  try {
-    const candidateList = toRerank.map(c =>
-      `[ID: ${c.id}] (${c.type}) ${c.title}${c.narrative ? ` — ${c.narrative.substring(0, 100)}` : ''}`,
-    ).join('\n');
-
-    const response = await callLLM(RERANK_PROMPT, `QUERY: ${query}\n\nCANDIDATES:\n${candidateList}`);
-
-    // Parse response — handle markdown code blocks
-    let content = response.content.trim();
-    if (content.startsWith('```')) {
-      content = content.replace(/^```(?:json)?\s*/, '').replace(/\s*```$/, '');
-    }
-
-    const rankedIds = JSON.parse(content) as string[];
-
-    // Validate: must be an array of IDs matching our candidates
-    if (!Array.isArray(rankedIds) || rankedIds.length === 0) {
-      return { reranked: candidates, usedLLM: true };
-    }
-
-    // Build reranked list preserving original scores for display
-    const idMap = new Map(toRerank.map(c => [c.id, c]));
-    const reranked: RerankCandidate[] = [];
-    const seen = new Set<string>();
-
-    // Add IDs in LLM-reranked order
-    for (const id of rankedIds) {
-      const candidate = idMap.get(id);
-      if (candidate && !seen.has(id)) {
-        reranked.push(candidate);
-        seen.add(id);
-      }
-    }
-
-    // Add any candidates the LLM missed (safety: never lose results)
-    for (const c of toRerank) {
-      if (!seen.has(c.id)) {
-        reranked.push(c);
-      }
-    }
-
-    // Append non-reranked tail
-    reranked.push(...rest);
-
-    return { reranked, usedLLM: true };
-  } catch {
-    return { reranked: candidates, usedLLM: false };
-  }
-}
-
-// ── Smart Compression Filtering ──────────────────────────────────
-
-/** Patterns that indicate already-concise content not worth compressing */
-const SKIP_PATTERNS = [
-  /^(?:Command|Run|Execute):\s/i,           // Shell commands
-  /^(?:File|Edit|Changed):\s/i,             // File change descriptions
-  /^git\s+(?:add|commit|push|pull|log)/i,   // Git operations
-  /^(?:npm|npx|pnpm|yarn|bun)\s/i,          // Package manager commands
-  /^(?:Remove-Item|New-Item|Set-Content)/i,  // PowerShell commands
-  /^[A-Za-z]:\\[\w\\]/,                      // Windows file paths
-  /^\/(?:usr|home|var|etc|opt)\//,           // Unix file paths
-];
-
-/** Low-value observation types that hooks auto-capture (usually already terse) */
-const LOW_COMPRESSION_TYPES = new Set(['what-changed', 'discovery', 'session-request']);
-
-/**
- * Determine if a narrative should skip LLM compression.
- *
- * Skip when:
- * - Content starts with command/path patterns (already structured, not prose)
- * - Type is hooks-auto-captured AND narrative is relatively short
- * - Narrative is mostly code/paths (high ratio of special chars)
- */
-function shouldSkipCompression(narrative: string, type?: string): boolean {
-  // Skip command/path-like content
-  const firstLine = narrative.split('\n')[0];
-  if (SKIP_PATTERNS.some(p => p.test(firstLine))) return true;
-
-  // Skip short auto-captured observations (hooks produce terse what-changed)
-  if (type && LOW_COMPRESSION_TYPES.has(type) && narrative.length < 200) return true;
-
-  // Skip if narrative is mostly code/structured data (high special char ratio)
-  const specialChars = (narrative.match(/[{}()\[\]<>:;=|\\\/\-_\.@#$%^&*+~`"']/g) || []).length;
-  if (specialChars / narrative.length > 0.35) return true;
-
-  return false;
-}
-
-// ── Utility ──────────────────────────────────────────────────────
-
-/** Rough token estimate: ~4 chars per token for English, ~2 for CJK */
-function estimateTokens(text: string): number {
-  const cjkChars = (text.match(/[\u4e00-\u9fff\u3040-\u309f\u30a0-\u30ff\uac00-\ud7af]/g) || []).length;
-  const otherChars = text.length - cjkChars;
-  return Math.ceil(cjkChars / 1.5 + otherChars / 4);
-}
+/**
+ * LLM Quality Enhancements
+ *
+ * Premium memory quality features powered by LLM:
+ * 1. Narrative Compression — compress verbose narratives into concise core knowledge
+ * 2. Search Reranking — rerank search results by relevance to current task context
+ *
+ * Both features gracefully degrade: when LLM is not configured, they return
+ * the original data unchanged.
+ *
+ * Performance targets:
+ * - Compression: ~60% token reduction per stored memory
+ * - Reranking: ~40% improvement in Top-5 precision
+ */
+
+import { callLLM, isLLMEnabled } from './provider.js';
+
+// ── Narrative Compression ────────────────────────────────────────
+
+const COMPRESS_PROMPT = `You are a memory compression engine for a coding assistant.
+
+Compress the given narrative while preserving ALL technical facts and reasoning.
+
+Rules:
+- Remove: filler words, debugging journey, repeated info already in facts
+- Keep: specific values, file paths, error messages, version numbers, config keys, causal relationships, design reasoning
+- Merge related points into dense sentences
+- If facts are provided separately, do NOT repeat them in the compressed narrative
+- Output the compressed text ONLY, no explanation or wrapper
+
+Examples:
+Input: "我在调试过程中发现JWT token的refresh机制存在问题，具体来说是因为服务端没有实现自动续签，导致用户在24小时后会遇到静默的认证失败，之前我一直以为是网络问题但后来排查发现是token过期了"
+Output: "JWT refresh无自动续签→24h后静默认证失败（非网络问题）"
+
+Input: "Final deployment model for shadcn-blog is stable: GitHub Actions build locally, SCP artifacts to VPS, systemd manages the process. Docker was considered but rejected due to complexity overhead for a simple blog. The whole pipeline takes about 2 minutes from push to live."
+Output: "shadcn-blog部署: GH Actions构建→SCP到VPS→systemd管理, 弃Docker(复杂度过高), push到上线~2min"`;
+
+/** Gentler prompt for high-value types where reasoning context matters */
+const COMPRESS_PROMPT_GENTLE = `You are a memory compression engine for a coding assistant.
+
+Lightly compress the given narrative — preserve reasoning, trade-offs, and "why" context.
+
+Rules:
+- Only remove: obvious filler, debugging detours, info already in the separate facts list
+- PRESERVE: design reasoning, rejected alternatives, trade-off analysis, causal chains
+- Aim for ~70-80% of original length, NOT aggressive compression
+- If facts are provided separately, do NOT repeat them in the compressed narrative
+- Output the compressed text ONLY, no explanation or wrapper`;
+
+/**
+ * Compress a narrative to its essential core using LLM.
+ *
+ * Returns the original narrative if:
+ * - LLM is not enabled
+ * - Narrative is already short (≤80 chars)
+ * - Narrative is already concise (commands, file paths, git operations)
+ * - LLM call fails
+ */
+export async function compressNarrative(
+  narrative: string,
+  facts?: string[],
+  type?: string,
+): Promise<{ compressed: string; saved: number; usedLLM: boolean }> {
+  const originalTokens = estimateTokens(narrative);
+
+  // Skip compression for short narratives (≤150 chars is already concise)
+  if (!isLLMEnabled() || narrative.length <= 150) {
+    return { compressed: narrative, saved: 0, usedLLM: false };
+  }
+
+  // Skip compression for already-concise content that LLM can't meaningfully compress
+  if (shouldSkipCompression(narrative, type)) {
+    return { compressed: narrative, saved: 0, usedLLM: false };
+  }
+
+  try {
+    const factsContext = facts && facts.length > 0
+      ? `\n\nSeparate facts (already stored, don't repeat): ${facts.join('; ')}`
+      : '';
+
+    // Use gentler compression for high-value types where reasoning matters
+    const HIGH_VALUE_TYPES = new Set(['decision', 'trade-off', 'why-it-exists', 'how-it-works']);
+    const prompt = (type && HIGH_VALUE_TYPES.has(type)) ? COMPRESS_PROMPT_GENTLE : COMPRESS_PROMPT;
+    const response = await callLLM(prompt, narrative + factsContext);
+    const compressed = response.content.trim();
+
+    // Sanity check: compressed should be shorter and non-empty
+    if (!compressed || compressed.length >= narrative.length) {
+      return { compressed: narrative, saved: 0, usedLLM: true };
+    }
+
+    const compressedTokens = estimateTokens(compressed);
+    return {
+      compressed,
+      saved: originalTokens - compressedTokens,
+      usedLLM: true,
+    };
+  } catch {
+    return { compressed: narrative, saved: 0, usedLLM: false };
+  }
+}
+
+// ── Search Reranking ─────────────────────────────────────────────
+
+/** Minimal search result for reranking */
+export interface RerankCandidate {
+  id: string;
+  title: string;
+  type: string;
+  score: number;
+  narrative?: string;
+}
+
+const RERANK_PROMPT = `You are a memory relevance ranker for a coding assistant.
+
+Given a QUERY (what the user/agent is looking for) and a list of CANDIDATE memories,
+rerank them by relevance to the query.
+
+Rules:
+- Consider semantic relevance, not just keyword overlap
+- Gotchas and decisions related to the query topic should rank higher
+- Recent problem-solutions for the same component should rank higher
+- Command or audit-log memories (titles starting with "Ran:" or "Command:") should rank lower for natural-language questions unless the query is explicitly about commands, scripts, or audit history
+- Generic or loosely related memories should rank lower
+- Output ONLY a JSON array of IDs in order of relevance (most relevant first)
+- Include ALL candidate IDs, just reorder them
+
+Example output: ["r1", "r3", "r2"]`;
+
+/**
+ * Rerank search results using LLM contextual understanding.
+ *
+ * Takes Orama's initial ranking and improves it by considering
+ * semantic relevance to the current query/task context.
+ *
+ * Returns original order if LLM is not enabled or call fails.
+ */
+export async function rerankResults(
+  query: string,
+  candidates: RerankCandidate[],
+): Promise<{ reranked: RerankCandidate[]; usedLLM: boolean }> {
+  // Skip if too few results or LLM not available
+  if (!isLLMEnabled() || candidates.length <= 2) {
+    return { reranked: candidates, usedLLM: false };
+  }
+
+  // Only rerank top-N to save LLM tokens (reranking 20+ is wasteful)
+  const MAX_RERANK = 10;
+  const toRerank = candidates.slice(0, MAX_RERANK);
+  const rest = candidates.slice(MAX_RERANK);
+
+  try {
+    const candidateList = toRerank.map(c =>
+      `[ID: ${c.id}] (${c.type}) ${c.title}${c.narrative ? ` — ${c.narrative.substring(0, 100)}` : ''}`,
+    ).join('\n');
+
+    const response = await callLLM(RERANK_PROMPT, `QUERY: ${query}\n\nCANDIDATES:\n${candidateList}`);
+
+    // Parse response — handle markdown code blocks
+    let content = response.content.trim();
+    if (content.startsWith('```')) {
+      content = content.replace(/^```(?:json)?\s*/, '').replace(/\s*```$/, '');
+    }
+
+    const rankedIds = JSON.parse(content) as string[];
+
+    // Validate: must be an array of IDs matching our candidates
+    if (!Array.isArray(rankedIds) || rankedIds.length === 0) {
+      return { reranked: candidates, usedLLM: true };
+    }
+
+    // Build reranked list preserving original scores for display
+    const idMap = new Map(toRerank.map(c => [c.id, c]));
+    const reranked: RerankCandidate[] = [];
+    const seen = new Set<string>();
+
+    // Add IDs in LLM-reranked order
+    for (const id of rankedIds) {
+      const candidate = idMap.get(id);
+      if (candidate && !seen.has(id)) {
+        reranked.push(candidate);
+        seen.add(id);
+      }
+    }
+
+    // Add any candidates the LLM missed (safety: never lose results)
+    for (const c of toRerank) {
+      if (!seen.has(c.id)) {
+        reranked.push(c);
+      }
+    }
+
+    // Append non-reranked tail
+    reranked.push(...rest);
+
+    return { reranked, usedLLM: true };
+  } catch {
+    return { reranked: candidates, usedLLM: false };
+  }
+}
+
+// ── Smart Compression Filtering ──────────────────────────────────
+
+/** Patterns that indicate already-concise content not worth compressing */
+const SKIP_PATTERNS = [
+  /^(?:Command|Run|Execute):\s/i,           // Shell commands
+  /^(?:File|Edit|Changed):\s/i,             // File change descriptions
+  /^git\s+(?:add|commit|push|pull|log)/i,   // Git operations
+  /^(?:npm|npx|pnpm|yarn|bun)\s/i,          // Package manager commands
+  /^(?:Remove-Item|New-Item|Set-Content)/i,  // PowerShell commands
+  /^[A-Za-z]:\\[\w\\]/,                      // Windows file paths
+  /^\/(?:usr|home|var|etc|opt)\//,           // Unix file paths
+];
+
+/** Low-value observation types that hooks auto-capture (usually already terse) */
+const LOW_COMPRESSION_TYPES = new Set(['what-changed', 'discovery', 'session-request']);
+
+/**
+ * Determine if a narrative should skip LLM compression.
+ *
+ * Skip when:
+ * - Content starts with command/path patterns (already structured, not prose)
+ * - Type is hooks-auto-captured AND narrative is relatively short
+ * - Narrative is mostly code/paths (high ratio of special chars)
+ */
+function shouldSkipCompression(narrative: string, type?: string): boolean {
+  // Skip command/path-like content
+  const firstLine = narrative.split('\n')[0];
+  if (SKIP_PATTERNS.some(p => p.test(firstLine))) return true;
+
+  // Skip short auto-captured observations (hooks produce terse what-changed)
+  if (type && LOW_COMPRESSION_TYPES.has(type) && narrative.length < 200) return true;
+
+  // Skip if narrative is mostly code/structured data (high special char ratio)
+  const specialChars = (narrative.match(/[{}()\[\]<>:;=|\\\/\-_\.@#$%^&*+~`"']/g) || []).length;
+  if (specialChars / narrative.length > 0.35) return true;
+
+  return false;
+}
+
+// ── Utility ──────────────────────────────────────────────────────
+
+/** Rough token estimate: ~4 chars per token for English, ~2 for CJK */
+function estimateTokens(text: string): number {
+  const cjkChars = (text.match(/[\u4e00-\u9fff\u3040-\u309f\u30a0-\u30ff\uac00-\ud7af]/g) || []).length;
+  const otherChars = text.length - cjkChars;
+  return Math.ceil(cjkChars / 1.5 + otherChars / 4);
+}