src-mcp 1.0.2 → 1.0.3

package/README.md

@@ -51,7 +51,6 @@ SRC indexes your codebase into semantic, searchable chunks that LLMs actually un
 | Feature | Description |
 |---------|-------------|
 | **Hybrid Search** | Vector + BM25 + RRF fusion for optimal results |
-| **LLM Re-ranking** | AI-powered relevance optimization |
 | **Call Graph** | Shows who calls what and what calls who |
 | **Cross-file Context** | Resolves imports and path aliases automatically |
 | **Incremental Updates** | SHA-256 hash detection for fast updates |
@@ -74,12 +73,11 @@ SRC indexes your codebase into semantic, searchable chunks that LLMs actually un
 
 ### 1. Install Ollama
 
-SRC requires [Ollama](https://ollama.com) for embeddings and re-ranking:
+SRC requires [Ollama](https://ollama.com) for embeddings:
 
 ```bash
 # Install from https://ollama.com, then:
 ollama pull nomic-embed-text
-ollama pull qwen2.5:1.5b
 ```
 
 ### 2. Install SRC
@@ -152,7 +150,6 @@ src-mcp get_index_status
 |------|----------|---------|-------------|
 | `search_code` | `--limit` | 10 | Max results |
 | `search_code` | `--mode` | hybrid | `hybrid` / `vector` / `fts` |
-| `search_code` | `--rerank` | true | LLM re-ranking |
 | `index_codebase` | `--concurrency` | 4 | Parallel workers |
 | `index_codebase` | `--force` | false | Re-index if exists |
 
@@ -234,7 +231,6 @@ Hybrid search with vector similarity, BM25 keyword matching, and RRF fusion.
 | `limit` | number | No | `10` | Maximum results to return |
 | `threshold` | number | No | — | Distance threshold (0-2, vector mode only) |
 | `mode` | enum | No | `hybrid` | Search mode: `hybrid`, `vector`, or `fts` |
-| `rerank` | boolean | No | `true` | Enable LLM re-ranking |
 | `includeCallContext` | boolean | No | `true` | Include caller/callee information |
 
 **Search Modes:**
@@ -405,7 +401,6 @@ All settings can be configured via environment variables:
 | `CHUNK_SIZE` | Characters per chunk | `1000` |
 | `CHUNK_OVERLAP` | Overlap between chunks | `200` |
 | `EMBEDDING_BATCH_SIZE` | Batch size for embedding | `10` |
-| `RERANK_MODEL` | Model for re-ranking | `qwen2.5:1.5b` |
 | `LOG_LEVEL` | Log verbosity | `info` |
 
 **Example:**
@@ -576,7 +571,7 @@ Source Files → Semantic Chunking → AST Enrichment → Cross-file Context →
 
 ```
 Query → Embed Query → Vector Search ─┐
-                                     ├→ RRF Fusion → LLM Re-rank → Add Call Context → Results
+                                     ├→ RRF Fusion → Add Call Context → Results
 Query → Tokenize ───→ BM25 Search ───┘
 ```
 
@@ -586,9 +581,8 @@ Query → Tokenize ───→ BM25 Search ───┘
 2. **Vector Search** — Find semantically similar chunks (cosine similarity)
 3. **BM25 Search** — Find keyword matches (term frequency)
 4. **RRF Fusion** — Combine rankings with Reciprocal Rank Fusion (k=60)
-5. **Re-rank** — Use LLM to re-order by relevance (qwen2.5:1.5b)
-6. **Call Context** — Add caller/callee information from call graph
-7. **Return** — Ranked results with full context
+5. **Call Context** — Add caller/callee information from call graph
+6. **Return** — Ranked results with full context
 
 ### Technical Specifications
 
@@ -596,7 +590,6 @@ Query → Tokenize ───→ BM25 Search ───┘
 |-----------|---------------|
 | **Embedding Model** | nomic-embed-text (137M params) |
 | **Vector Dimensions** | 768 |
-| **Re-ranking Model** | qwen2.5:1.5b (1.5B params) |
 | **Chunk Size** | 1000 characters |
 | **Chunk Overlap** | 200 characters |
 | **Batch Size** | 10 embeddings per request |
@@ -612,7 +605,6 @@ Query → Tokenize ───→ BM25 Search ───┘
 | Feature | SRC | Basic MCPs |
 |---------|-----|------------|
 | **Search Method** | Hybrid (Vector + BM25 + RRF) | Keyword only or basic embedding |
-| **Re-ranking** | LLM-powered | None |
 | **Call Graph** | Full caller/callee context | None |
 | **Cross-file Context** | Resolves imports & path aliases | None |
 | **Incremental Updates** | SHA-256 hash detection | Full re-index required |
@@ -622,11 +614,10 @@ Query → Tokenize ───→ BM25 Search ───┘
 ### Key Advantages
 
 1. **Hybrid Search** — Combines semantic understanding with keyword precision
-2. **LLM Re-ranking** — Catches nuances that pure vector search misses
-3. **Call Graph** — Understand code relationships, not just content
-4. **Cross-file Resolution** — Follows imports to provide complete context
-5. **Incremental Updates** — Only re-index what changed
-6. **Semantic Chunking** — Splits at symbol boundaries, not arbitrary lines
+2. **Call Graph** — Understand code relationships, not just content
+3. **Cross-file Resolution** — Follows imports to provide complete context
+4. **Incremental Updates** — Only re-index what changed
+5. **Semantic Chunking** — Splits at symbol boundaries, not arbitrary lines
 
 ---
 
@@ -652,7 +643,6 @@ Error: model 'nomic-embed-text' not found
 **Solution:**
 ```bash
 ollama pull nomic-embed-text
-ollama pull qwen2.5:1.5b
 ```
 
 ### Index Already Exists

package/dist/bin.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { a as logger, i as colors, n as features, o as EMBEDDING_CONFIG, r as createIndexWatcher, s as config, t as startServer } from "./server-B2Ms4jQx.mjs";
+import { a as logger, i as colors, n as features, o as EMBEDDING_CONFIG, r as createIndexWatcher, s as config, t as startServer } from "./server-DL8hfycz.mjs";
 import { defineCommand, runMain } from "citty";
 
 //#region src/cli/parser.ts

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { t as startServer } from "./server-B2Ms4jQx.mjs";
+import { t as startServer } from "./server-DL8hfycz.mjs";
 
 //#region src/index.ts
 startServer().catch(console.error);

package/dist/{server-B2Ms4jQx.mjs → server-DL8hfycz.mjs}

@@ -21,7 +21,7 @@ import "@langchain/textsplitters";
 const config = {
 	name: "src-mcp",
 	fullName: "SRC (Structured Repo Context)",
-	version: "1.0.2",
+	version: "1.0.3",
 	description: "MCP server for codebase analysis with Treesitter (SCM queries), AST parsing, and embedding-based indexing"
 };
 const nodeEnv = process.env.NODE_ENV;
@@ -40,8 +40,7 @@ const EMBEDDING_CONFIG = {
 	embeddingDimensions: Number(process.env.EMBEDDING_DIMENSIONS) || 768,
 	defaultChunkSize: Number(process.env.CHUNK_SIZE) || 1e3,
 	defaultChunkOverlap: Number(process.env.CHUNK_OVERLAP) || 200,
-	batchSize: Number(process.env.EMBEDDING_BATCH_SIZE) || 10,
-	rerankModel: process.env.RERANK_MODEL ?? "qwen2.5:1.5b"
+	batchSize: Number(process.env.EMBEDDING_BATCH_SIZE) || 10
 };
 /**
 * Enrichment configuration for cross-file context
@@ -2387,8 +2386,9 @@ var IndexWatcher = class {
 	async start() {
 		const health = await this.ollamaClient.healthCheck();
 		if (!health.ok) throw new Error(health.error ?? "Ollama is not available");
+		const needsFullIndex = !this.vectorStore.exists();
 		await this.vectorStore.connect();
-		if (!this.vectorStore.exists()) await this.fullIndex();
+		if (needsFullIndex) await this.fullIndex();
 		this.watcher = watch(this.directory, {
 			ignored: (filePath) => {
 				const relativePath = path.relative(this.directory, filePath).replace(/\\/g, "/");
@@ -2468,96 +2468,6 @@ function createIndexWatcher(options) {
 	return new IndexWatcher(options);
 }
 
-//#endregion
-//#region src/core/embeddings/reranker.ts
-/**
-* Parse LLM response to extract relevance score
-*/
-function parseScore(response) {
-	const match = /\b(\d+(?:\.\d+)?)\b/.exec(response);
-	if (match?.[1]) {
-		const score = parseFloat(match[1]);
-		if (score >= 0 && score <= 10) return score;
-		if (score > 10 && score <= 100) return score / 10;
-	}
-	return 5;
-}
-/**
-* Score a single query-document pair using Ollama
-*/
-async function scoreResult(query, content, options) {
-	const model = options.model ?? "llama3.2";
-	const timeout = options.timeout ?? 3e4;
-	const prompt = `Rate the relevance of the following code snippet to the search query on a scale of 0-10.
-0 = completely irrelevant
-5 = somewhat relevant
-10 = highly relevant and directly answers the query
-
-Query: "${query}"
-
-Code:
-\`\`\`
-${content.slice(0, 1e3)}
-\`\`\`
-
-Respond with ONLY a number between 0 and 10.`;
-	try {
-		const response = await fetch(`${options.ollamaBaseUrl}/api/generate`, {
-			method: "POST",
-			headers: { "Content-Type": "application/json" },
-			body: JSON.stringify({
-				model,
-				prompt,
-				stream: false,
-				options: {
-					temperature: 0,
-					num_predict: 10
-				}
-			}),
-			signal: AbortSignal.timeout(timeout)
-		});
-		if (!response.ok) {
-			logger.warn(`Re-ranking request failed: ${response.statusText}`);
-			return 5;
-		}
-		return parseScore((await response.json()).response ?? "5");
-	} catch (error) {
-		logger.warn(`Re-ranking error: ${error instanceof Error ? error.message : String(error)}`);
-		return 5;
-	}
-}
-/**
-* Re-rank search results using LLM scoring
-*
-* Takes initial search results and re-scores them based on
-* semantic relevance to the query using an LLM.
-*/
-async function rerank(query, results, options) {
-	const maxResults = options.maxResults ?? 20;
-	const toRerank = results.slice(0, maxResults);
-	if (toRerank.length === 0) return [];
-	logger.debug(`Re-ranking ${String(toRerank.length)} results for: ${query}`);
-	const batchSize = 5;
-	const rerankedResults = [];
-	for (let i = 0; i < toRerank.length; i += batchSize) {
-		const batch = toRerank.slice(i, i + batchSize);
-		const scores = await Promise.all(batch.map(async (result) => scoreResult(query, result.chunk.content, options)));
-		for (let j = 0; j < batch.length; j++) {
-			const result = batch[j];
-			const score = scores[j];
-			if (result !== void 0 && score !== void 0) rerankedResults.push({
-				...result,
-				originalScore: result.score,
-				rerankScore: score,
-				score
-			});
-		}
-	}
-	rerankedResults.sort((a, b) => b.rerankScore - a.rerankScore);
-	logger.debug(`Re-ranking complete, top score: ${String(rerankedResults[0]?.rerankScore ?? 0)}`);
-	return rerankedResults;
-}
-
 //#endregion
 //#region src/core/embeddings/callgraph.ts
 /**
@@ -3051,7 +2961,6 @@ const searchCodeSchema = z.object({
 		"fts",
 		"hybrid"
 	]).optional().default("hybrid").describe("Search mode: 'vector' (semantic only), 'fts' (keyword only), 'hybrid' (combined with RRF fusion)"),
-	rerank: z.boolean().optional().default(true).describe("Enable LLM re-ranking for improved relevance (enabled by default)"),
 	includeCallContext: z.boolean().optional().default(true).describe("Include caller/callee information for each result (uses cached call graph)")
 });
 /**
@@ -3114,7 +3023,7 @@ function formatResults(results, baseDir) {
 * Execute the search_code feature
 */
 async function execute$2(input) {
-	const { query, directory, limit, threshold, mode, rerank: enableRerank, includeCallContext } = input;
+	const { query, directory, limit, threshold, mode, includeCallContext } = input;
 	if (!fs.existsSync(directory)) return {
 		success: false,
 		error: `Directory not found: ${directory}`
@@ -3136,11 +3045,6 @@ async function execute$2(input) {
 		const queryVector = await ollamaClient.embed(query);
 		let results = await vectorStore.searchHybrid(queryVector, query, limit, { mode });
 		if (threshold !== void 0 && mode === "vector") results = results.filter((r) => r.score <= threshold);
-		if (enableRerank && results.length > 0) results = await rerank(query, results, {
-			ollamaBaseUrl: EMBEDDING_CONFIG.ollamaBaseUrl,
-			model: EMBEDDING_CONFIG.rerankModel,
-			maxResults: limit
-		});
 		vectorStore.close();
 		let formattedResults = formatResults(results, absoluteDir);
 		if (includeCallContext && formattedResults.length > 0) {
@@ -3713,7 +3617,6 @@ SRC is a semantic code search MCP server. It indexes codebases and provides inte
 - **Vector embeddings** for semantic similarity (understands meaning, not just keywords)
 - **BM25 keyword search** for exact matches
 - **Hybrid search** combining both with RRF fusion
-- **LLM re-ranking** for optimal relevance
 - **Call graph analysis** showing function relationships
 
 ## When to use SRC?
@@ -3781,7 +3684,6 @@ search_code(query: "your search query here")
 | query | string | required | Natural language search query |
 | limit | number | 10 | Max results to return |
 | mode | "hybrid" / "vector" / "fts" | "hybrid" | Search mode |
-| rerank | boolean | true | LLM re-ranking for better relevance |
 | includeCallContext | boolean | true | Include caller/callee info |
 | threshold | number | - | Distance threshold (vector mode only) |
 
@@ -3886,4 +3788,4 @@ async function startServer() {
 
 //#endregion
 export { logger as a, colors as i, features as n, EMBEDDING_CONFIG as o, createIndexWatcher as r, config as s, startServer as t };
-//# sourceMappingURL=server-B2Ms4jQx.mjs.map
+//# sourceMappingURL=server-DL8hfycz.mjs.map