feat: use fine-tuned 1.7B expansion model with chunked external data

Split the 2.32GB Q4 data file into 2 chunks (~1.16GB each) to stay
under the 2GB protobuf limit. Transformers.js v4 supports multi-chunk
external data via config.json use_external_data_format field.

Restored original fine-tuned prompt format (/no_think Expand this...).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/constants.ts

@@ -33,7 +33,7 @@ export const EMBED_DOC_TEMPLATE = (title: string, body: string) =>
 // Model IDs for Transformers.js
 export const MODEL_EMBEDDING = "shreyask/embeddinggemma-300m-ONNX";
 export const MODEL_RERANKER = "onnx-community/Qwen3-Reranker-0.6B-ONNX";
-export const MODEL_EXPANSION = "onnx-community/Qwen3-0.6B-ONNX";
+export const MODEL_EXPANSION = "shreyask/qmd-query-expansion-1.7B-ONNX";
 
 // Example queries
 export const EXAMPLE_QUERIES = [

src/pipeline/expansion.ts

@@ -1,14 +1,9 @@
 import { getExpansionPipeline } from "./models";
 import type { ExpandedQuery } from "../types";
 
-// The expansion model prompt format
-// For the fine-tuned model: `/no_think Expand this search query: ${query}`
-// For generic Qwen3-0.6B: use a structured instruction prompt
+// The fine-tuned expansion model prompt format (from QMD)
 const EXPANSION_PROMPT = (query: string) =>
-  `/no_think Given the search query "${query}", generate three types of expanded queries for a hybrid search system. Output exactly this format with no extra text:
-lex: keyword1 keyword2 keyword3 keyword4 keyword5
-vec: A natural sentence that captures the semantic meaning of the query
-hyde: A short paragraph that a relevant document might contain about this topic`;
+  `/no_think Expand this search query: ${query}`;
 
 // Parse the model's output text into structured ExpandedQuery
 function parseExpansionOutput(text: string): ExpandedQuery {