kontext-engine 0.1.2 → 0.1.3

package/dist/cli/index.js

@@ -1208,12 +1208,15 @@ function createDatabase(dbPath, dimensions = DEFAULT_DIMENSIONS) {
       const rows = db.prepare(
         `SELECT c.id, c.file_id as fileId, f.path as filePath, f.language,
                   c.line_start as lineStart, c.line_end as lineEnd,
-                  c.type, c.name, c.parent, c.text
+                  c.type, c.name, c.parent, c.text, c.exports as exports
            FROM chunks c
            JOIN files f ON f.id = c.file_id
            WHERE c.id IN (${placeholders})`
       ).all(...ids);
-      return rows;
+      return rows.map((r) => ({
+        ...r,
+        exports: r.exports === 1
+      }));
     },
     searchChunks(filters, limit) {
       const conditions = [];
@@ -1250,7 +1253,7 @@ function createDatabase(dbPath, dimensions = DEFAULT_DIMENSIONS) {
       const sql = `
         SELECT c.id, c.file_id as fileId, f.path as filePath, f.language,
                c.line_start as lineStart, c.line_end as lineEnd,
-               c.type, c.name, c.parent, c.text
+               c.type, c.name, c.parent, c.text, c.exports as exports
         FROM chunks c
         JOIN files f ON f.id = c.file_id
         ${where}
@@ -1258,7 +1261,11 @@ function createDatabase(dbPath, dimensions = DEFAULT_DIMENSIONS) {
         LIMIT ?
       `;
       params.push(limit);
-      return db.prepare(sql).all(...params);
+      const rows = db.prepare(sql).all(...params);
+      return rows.map((r) => ({
+        ...r,
+        exports: r.exports === 1
+      }));
     },
     deleteChunksByFile(fileId) {
       const chunkRows = stmtGetChunkIdsByFile.all(fileId);
@@ -1558,6 +1565,7 @@ async function vectorSearch(db, embedder, query, limit, filters) {
       lineEnd: chunk.lineEnd,
       name: chunk.name,
       type: chunk.type,
+      exported: chunk.exports,
       text: chunk.text,
       score: distanceToScore(vr.distance),
       language: chunk.language
@@ -1598,6 +1606,7 @@ function ftsSearch(db, query, limit, filters) {
       lineEnd: chunk.lineEnd,
       name: chunk.name,
       type: chunk.type,
+      exported: chunk.exports,
       text: chunk.text,
       score: bm25ToScore(fts.rank),
       language: chunk.language
@@ -1632,6 +1641,7 @@ function astSearch(db, filters, limit) {
     lineEnd: chunk.lineEnd,
     name: chunk.name,
     type: chunk.type,
+    exported: chunk.exports,
     text: chunk.text,
     score,
     language: chunk.language
@@ -1684,6 +1694,7 @@ function pathSearch(db, pattern, limit) {
         lineEnd: chunk.lineEnd,
         name: chunk.name,
         type: chunk.type,
+        exported: chunk.exports,
         text: chunk.text,
         score: 1,
         language: file.language
@@ -1726,6 +1737,7 @@ function pathKeywordSearch(db, query, limit) {
         lineEnd: chunk.lineEnd,
         name: chunk.name,
         type: chunk.type,
+        exported: chunk.exports,
         text: chunk.text,
         score,
         language: file.language
@@ -1782,11 +1794,24 @@ var PATH_BOOST_DIR_EXACT = 1.5;
 var PATH_BOOST_FILENAME = 1.4;
 var PATH_BOOST_PARTIAL = 1.2;
 var IMPORT_PENALTY = 0.5;
+var TEST_FILE_PENALTY = 0.65;
+var SMALL_SNIPPET_PENALTY = 0.75;
+var PUBLIC_API_BOOST = 1.12;
+var TEST_FILE_DIRECTORY_PATTERN = /(?:^|\/)(?:tests|__tests__)(?:\/|$)/;
+var TEST_FILE_NAME_PATTERN = /(?:^|\/)[^/]*\.(?:test|spec)\.[cm]?[jt]sx?$/;
+var SMALL_SNIPPET_MAX_LINES = 3;
+function extractPathBoostTerms(query) {
+  return query.split(/\s+/).map((t) => t.trim()).filter((t) => t.length >= 2);
+}
 function fusionMergeWithPathBoost(strategyResults, limit, pathBoostTerms) {
   const fused = fusionMerge(strategyResults, limit * 3);
   if (fused.length === 0) return [];
   const boosted = applyPathBoost(fused, pathBoostTerms);
-  const adjusted = applyImportDeprioritization(boosted);
+  const importAdjusted = applyImportDeprioritization(boosted);
+  const testAdjusted = applyTestFileDeprioritization(importAdjusted);
+  const snippetAdjusted = applySmallSnippetDeprioritization(testAdjusted);
+  const boostedApi = applyPublicApiBoost(snippetAdjusted);
+  const adjusted = applyFileDiversityDiminishingReturns(boostedApi);
   adjusted.sort((a, b) => b.score - a.score);
   const sliced = adjusted.slice(0, limit);
   return renormalize(sliced);
@@ -1836,6 +1861,76 @@ function applyImportDeprioritization(results) {
     return r;
   });
 }
+function applyTestFileDeprioritization(results) {
+  const hasNonTestFile = results.some((r) => !isTestFilePath(r.filePath));
+  if (!hasNonTestFile) return results;
+  const maxNonTestScore = Math.max(
+    ...results.filter((r) => !isTestFilePath(r.filePath)).map((r) => r.score),
+    0
+  );
+  if (maxNonTestScore === 0) return results;
+  return results.map((r) => {
+    if (isTestFilePath(r.filePath)) {
+      return { ...r, score: r.score * TEST_FILE_PENALTY };
+    }
+    return r;
+  });
+}
+function applySmallSnippetDeprioritization(results) {
+  const hasNonSmallSnippet = results.some((r) => !isSmallSnippet(r));
+  if (!hasNonSmallSnippet) return results;
+  const maxNonSmallScore = Math.max(
+    ...results.filter((r) => !isSmallSnippet(r)).map((r) => r.score),
+    0
+  );
+  if (maxNonSmallScore === 0) return results;
+  return results.map((r) => {
+    if (isSmallSnippet(r)) {
+      return { ...r, score: r.score * SMALL_SNIPPET_PENALTY };
+    }
+    return r;
+  });
+}
+function applyPublicApiBoost(results) {
+  return results.map((r) => {
+    if (isPublicApiSymbol(r)) {
+      return { ...r, score: r.score * PUBLIC_API_BOOST };
+    }
+    return r;
+  });
+}
+function applyFileDiversityDiminishingReturns(results) {
+  if (results.length <= 1) return results;
+  const ranked = [...results].sort((a, b) => b.score - a.score);
+  const seenPerFile = /* @__PURE__ */ new Map();
+  return ranked.map((r) => {
+    const count = (seenPerFile.get(r.filePath) ?? 0) + 1;
+    seenPerFile.set(r.filePath, count);
+    return {
+      ...r,
+      score: r.score * getFileDiversityFactor(count)
+    };
+  });
+}
+function isTestFilePath(filePath) {
+  const normalizedPath = filePath.toLowerCase().replace(/\\/g, "/");
+  return TEST_FILE_DIRECTORY_PATTERN.test(normalizedPath) || TEST_FILE_NAME_PATTERN.test(normalizedPath);
+}
+function isSmallSnippet(result) {
+  const lineCount = Math.max(1, result.lineEnd - result.lineStart + 1);
+  return lineCount <= SMALL_SNIPPET_MAX_LINES;
+}
+function isPublicApiSymbol(result) {
+  if (result.exported === true) return true;
+  const textStart = result.text.trimStart().toLowerCase();
+  return textStart.startsWith("export ");
+}
+function getFileDiversityFactor(fileOccurrence) {
+  if (fileOccurrence <= 1) return 1;
+  if (fileOccurrence === 2) return 0.9;
+  if (fileOccurrence === 3) return 0.8;
+  return 0.7;
+}
 function renormalize(results) {
   if (results.length === 0) return results;
   const maxScore = Math.max(...results.map((r) => r.score));
@@ -1896,9 +1991,6 @@ function extractSymbolNames(query) {
 function isPathLike(query) {
   return query.includes("/") || query.includes("*") || query.includes(".");
 }
-function extractPathBoostTerms(query) {
-  return query.split(/\s+/).map((t) => t.trim()).filter((t) => t.length >= 2);
-}
 async function runQuery(projectPath, query, options) {
   const absoluteRoot = path5.resolve(projectPath);
   const dbPath = path5.join(absoluteRoot, CTX_DIR2, DB_FILENAME2);
@@ -2028,27 +2120,128 @@ function registerQueryCommand(program2) {
 import fs7 from "fs";
 import path6 from "path";
 
+// src/steering/prompts.ts
+var PLAN_SYSTEM_PROMPT = `You are a code-search strategy planner for a TypeScript/JavaScript codebase.
+
+Given a user query, produce a JSON object with:
+- "interpretation": one sentence summarising what the user wants to find.
+- "strategies": an ordered array of search strategies (most important first).
+
+Each strategy object has:
+  "strategy" \u2014 one of "vector", "fts", "ast", "path", "dependency"
+  "query"    \u2014 the optimised search string for that strategy (see rules below)
+  "weight"   \u2014 importance 0\u20131 (highest-priority strategy gets 1.0)
+  "reason"   \u2014 one sentence explaining why this strategy helps
+
+## Strategy selection rules
+
+| Signal in query | Primary strategy | Supporting strategies |
+|---|---|---|
+| Conceptual / "how does X work" / natural language | vector | fts, ast |
+| Exact keyword, identifier, or error message | fts | ast |
+| Symbol name (function, class, type, variable) | ast | fts |
+| File path, glob, or extension (e.g. "*.test.ts") | path | fts |
+| Import chain / "what depends on X" | dependency | ast, fts |
+| Mixed: natural language + code symbol | vector + ast | fts |
+
+## Query optimisation rules
+- **vector**: keep the query close to natural language; rephrase for semantic similarity.
+- **fts**: extract the most distinctive keywords/identifiers; drop stop words.
+- **ast**: use only the symbol name (camelCase, snake_case, or PascalCase). Strip surrounding prose.
+- **path**: use a glob or slash-separated path segment (e.g. "src/auth/*.ts").
+- **dependency**: use the bare module or file name being imported.
+
+## Edge cases
+- **Vague query** (e.g. "help me understand this"): use vector with the full query; add fts with any nouns present.
+- **Multi-concept query** (e.g. "authentication and rate limiting"): create separate strategies for each concept, both at high weight.
+- **Code symbol mixed with prose** (e.g. "where is the validateToken function called"): use ast for the symbol and vector for the intent.
+- **Query is just a symbol** (e.g. "createPool"): use ast at weight 1.0 and fts at weight 0.7. Skip vector.
+
+## Examples
+
+User: "how does authentication work"
+\`\`\`json
+{
+  "interpretation": "Understand the authentication flow and related middleware.",
+  "strategies": [
+    { "strategy": "vector", "query": "authentication flow middleware", "weight": 1.0, "reason": "Conceptual question best served by semantic search." },
+    { "strategy": "fts", "query": "authentication middleware auth", "weight": 0.7, "reason": "Keyword fallback for auth-related identifiers." },
+    { "strategy": "ast", "query": "authenticate", "weight": 0.6, "reason": "Likely function or class name." }
+  ]
+}
+\`\`\`
+
+User: "validateToken"
+\`\`\`json
+{
+  "interpretation": "Find the validateToken symbol definition and usages.",
+  "strategies": [
+    { "strategy": "ast", "query": "validateToken", "weight": 1.0, "reason": "Exact symbol lookup." },
+    { "strategy": "fts", "query": "validateToken", "weight": 0.7, "reason": "Catch references in comments or strings." }
+  ]
+}
+\`\`\`
+
+User: "where is rate limiting configured in src/middleware"
+\`\`\`json
+{
+  "interpretation": "Locate rate-limiting configuration inside the middleware directory.",
+  "strategies": [
+    { "strategy": "path", "query": "src/middleware/*", "weight": 0.9, "reason": "Scope results to the specified directory." },
+    { "strategy": "vector", "query": "rate limiting configuration", "weight": 1.0, "reason": "Semantic match for the concept." },
+    { "strategy": "fts", "query": "rateLimit rateLimiter", "weight": 0.7, "reason": "Common identifier variants." }
+  ]
+}
+\`\`\`
+
+User: "authentication and database connection pooling"
+\`\`\`json
+{
+  "interpretation": "Find code related to both authentication and database connection pooling.",
+  "strategies": [
+    { "strategy": "vector", "query": "authentication login", "weight": 1.0, "reason": "Semantic search for the auth concept." },
+    { "strategy": "vector", "query": "database connection pool", "weight": 1.0, "reason": "Semantic search for the DB pooling concept." },
+    { "strategy": "fts", "query": "auth createPool connectionPool", "weight": 0.7, "reason": "Keyword fallback for likely identifiers." }
+  ]
+}
+\`\`\`
+
+Output ONLY the JSON object. No markdown fences, no commentary.`;
+var SYNTHESIZE_SYSTEM_PROMPT = `You are a code-search assistant. Given a user query and ranked search results, produce a concise, actionable summary.
+
+## Output structure (plain text, no markdown)
+
+1. **Key finding** (1\u20132 sentences): the most important result or answer first.
+2. **Supporting locations** (bulleted, max 5): each line is "filePath:lineStart \u2013 brief description".
+3. **Additional context** (0\u20132 sentences, optional): relationships between results, patterns, or next steps.
+
+## Rules
+- Always reference file paths and line numbers from the search results.
+- Mention specific symbol names (functions, classes, types) when they appear in results.
+- If no result clearly answers the query, say so and suggest a refined search.
+- Be concise \u2014 aim for 4\u20138 lines total. Do not repeat the query back.
+- Do not use markdown formatting (no #, *, \`, or fences). Use plain text only.
+- Group related results rather than listing every result individually.
+
+## Example
+
+Query: "how does token validation work"
+Results include validateToken in src/auth/tokens.ts:42 and authMiddleware in src/middleware/auth.ts:15.
+
+Good output:
+Token validation is handled by validateToken (src/auth/tokens.ts:42), which decodes a JWT and checks expiry and signature against the configured secret.
+
+Related locations:
+- src/auth/tokens.ts:42 \u2013 validateToken: core JWT decode + verify logic
+- src/middleware/auth.ts:15 \u2013 authMiddleware: calls validateToken on every protected route
+- src/auth/types.ts:5 \u2013 TokenPayload type definition
+
+The middleware extracts the Bearer token from the Authorization header before passing it to validateToken.`;
+
 // src/steering/llm.ts
 var GEMINI_URL = "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-flash-preview:generateContent";
 var OPENAI_URL = "https://api.openai.com/v1/responses";
 var ANTHROPIC_URL = "https://api.anthropic.com/v1/messages";
-var PLAN_SYSTEM_PROMPT = `You are a code search strategy planner. Given a user query about code, output a JSON object with:
-- "interpretation": a one-line summary of what the user is looking for
-- "strategies": an array of search strategy objects, each with:
-  - "strategy": one of "vector", "fts", "ast", "path", "dependency"
-  - "query": the optimized query string for that strategy
-  - "weight": a number 0-1 indicating importance
-  - "reason": brief explanation of why this strategy is used
-
-Choose strategies based on query type:
-- Conceptual/natural language \u2192 vector (semantic search)
-- Keywords/identifiers \u2192 fts (full-text search)
-- Symbol names (functions, classes) \u2192 ast (structural search)
-- File paths or patterns \u2192 path (path glob search)
-- Import/dependency chains \u2192 dependency
-
-Output ONLY valid JSON, no markdown.`;
-var SYNTHESIZE_SYSTEM_PROMPT = `You are a code search assistant. Given search results, write a brief, helpful explanation of what was found. Be concise (2-4 sentences). Reference specific files and function names. Do not use markdown.`;
 function createGeminiProvider(apiKey) {
   return {
     name: "gemini",
@@ -2156,6 +2349,7 @@ function createAnthropicProvider(apiKey) {
   };
 }
 var STOP_WORDS = /* @__PURE__ */ new Set([
+  // Interrogatives & conjunctions
   "how",
   "does",
   "what",
@@ -2165,6 +2359,7 @@ var STOP_WORDS = /* @__PURE__ */ new Set([
   "which",
   "who",
   "whom",
+  // Be-verbs
   "is",
   "are",
   "was",
@@ -2172,10 +2367,12 @@ var STOP_WORDS = /* @__PURE__ */ new Set([
   "be",
   "been",
   "being",
+  // Do-verbs
   "do",
   "did",
   "doing",
   "done",
+  // Articles, connectors, prepositions
   "the",
   "a",
   "an",
@@ -2194,12 +2391,30 @@ var STOP_WORDS = /* @__PURE__ */ new Set([
   "by",
   "from",
   "about",
+  "into",
+  "through",
+  "between",
+  "after",
+  "before",
+  "during",
+  // Pronouns & demonstratives
   "it",
   "its",
   "this",
   "that",
   "these",
   "those",
+  "i",
+  "me",
+  "my",
+  "we",
+  "our",
+  "you",
+  "your",
+  "he",
+  "she",
+  "they",
+  // Modals
   "can",
   "could",
   "should",
@@ -2208,32 +2423,69 @@ var STOP_WORDS = /* @__PURE__ */ new Set([
   "shall",
   "may",
   "might",
+  // Have-verbs
   "has",
   "have",
   "had",
   "having",
-  "i",
-  "me",
-  "my",
-  "we",
-  "our",
-  "you",
-  "your",
-  "he",
-  "she",
-  "they",
+  // Common imperative verbs that carry no search value
   "find",
   "show",
   "get",
-  "tell"
+  "tell",
+  "look",
+  "give",
+  "list",
+  "explain",
+  // Misc filler
+  "all",
+  "any",
+  "some",
+  "each",
+  "every",
+  "much",
+  "many",
+  "also",
+  "just",
+  "like",
+  "then",
+  "there",
+  "here",
+  "very",
+  "really",
+  "use",
+  "used",
+  "using"
 ]);
+var CODE_IDENT_RE = /^(?:[a-z]+(?:[A-Z][a-z]*)+|[A-Z][a-zA-Z]+|[a-z]+(?:_[a-z]+)+|[A-Z]+(?:_[A-Z]+)+)$/;
+var DOTTED_IDENT_RE = /[a-zA-Z_]\w*(?:\.[a-zA-Z_]\w*)+/g;
 function extractSearchTerms(query) {
-  const words = query.replace(/[^a-zA-Z0-9_\s]/g, " ").split(/\s+/).filter((w) => w.length >= 2 && !STOP_WORDS.has(w.toLowerCase()));
-  if (words.length === 0) {
+  const terms = [];
+  const seen = /* @__PURE__ */ new Set();
+  const addUnique = (term) => {
+    const key = term.toLowerCase();
+    if (!seen.has(key)) {
+      seen.add(key);
+      terms.push(term);
+    }
+  };
+  const dottedMatches = query.match(DOTTED_IDENT_RE) ?? [];
+  for (const m of dottedMatches) addUnique(m);
+  const pathTokens = query.split(/\s+/).filter((t) => t.includes("/"));
+  for (const p of pathTokens) addUnique(p.replace(/[?!,;]+$/g, ""));
+  const words = query.replace(/[^a-zA-Z0-9_.\s/-]/g, " ").split(/\s+/).filter((w) => w.length >= 2);
+  for (const w of words) {
+    const lower = w.toLowerCase();
+    if (seen.has(lower)) continue;
+    if (STOP_WORDS.has(lower) && !CODE_IDENT_RE.test(w)) continue;
+    addUnique(w);
+  }
+  if (terms.length === 0) {
     const allWords = query.replace(/[^a-zA-Z0-9_\s]/g, " ").split(/\s+/).filter((w) => w.length >= 2);
-    return allWords.sort((a, b) => b.length - a.length)[0] ?? query;
+    const longest = allWords.sort((a, b) => b.length - a.length)[0];
+    return longest ?? query;
   }
-  return words.join(" ");
+  return terms.join(" ");
 }
 var VALID_STRATEGIES = /* @__PURE__ */ new Set([
   "vector",
@@ -2418,7 +2670,8 @@ function formatTextOutput2(output) {
   );
   return lines.join("\n");
 }
-function createSearchExecutor(db) {
+function createSearchExecutor(db, query) {
+  const pathBoostTerms = extractPathBoostTerms(query);
   return async (strategies, limit) => {
     const strategyResults = [];
     const fetchLimit = limit * 3;
@@ -2432,7 +2685,7 @@ function createSearchExecutor(db) {
         });
       }
     }
-    return fusionMerge(strategyResults, limit);
+    return fusionMergeWithPathBoost(strategyResults, limit, pathBoostTerms);
   };
 }
 function extractSymbolNames2(query) {
@@ -2480,7 +2733,7 @@ async function loadEmbedder2() {
   return embedderInstance2;
 }
 async function fallbackSearch(db, query, limit) {
-  const executor = createSearchExecutor(db);
+  const executor = createSearchExecutor(db, query);
   const keywords = extractSearchTerms(query);
   const fallbackStrategies = [
     { strategy: "fts", query: keywords, weight: 0.8, reason: "fallback keyword search" },
@@ -2522,7 +2775,7 @@ async function runAsk(projectPath, query, options) {
       }
       return output;
     }
-    const executor = createSearchExecutor(db);
+    const executor = createSearchExecutor(db, query);
     if (options.noExplain) {
       return await runNoExplain(provider, query, options, executor);
     }