Make embedding model settings more future-proof (#454)

src/lib/server/websearch/runWebSearch.ts

@@ -4,7 +4,10 @@ import type { WebSearch, WebSearchSource } from "$lib/types/WebSearch";
 import { generateQuery } from "$lib/server/websearch/generateQuery";
 import { parseWeb } from "$lib/server/websearch/parseWeb";
 import { chunk } from "$lib/utils/chunk";
-import { findSimilarSentences } from "$lib/server/websearch/sentenceSimilarity";
+import {
+	MAX_SEQ_LEN as CHUNK_CAR_LEN,
+	findSimilarSentences,
+} from "$lib/server/websearch/sentenceSimilarity";
 import type { Conversation } from "$lib/types/Conversation";
 import type { MessageUpdate } from "$lib/types/MessageUpdate";
 
@@ -62,7 +65,6 @@ export async function runWebSearch(
 				} catch (e) {
 					console.error(`Error parsing webpage "${link}"`, e);
 				}
-				const CHUNK_CAR_LEN = 512;
 				const MAX_N_CHUNKS = 100;
 				const texts = chunk(text, CHUNK_CAR_LEN).slice(0, MAX_N_CHUNKS);
 				return texts.map((t) => ({ source: result, text: t }));

src/lib/server/websearch/sentenceSimilarity.ts

@@ -6,7 +6,10 @@ function innerProduct(tensor1: Tensor, tensor2: Tensor) {
 	return 1.0 - dot(tensor1.data, tensor2.data);
 }
 
-const extractor = await pipeline("feature-extraction", "Xenova/gte-small");
+const modelId = "Xenova/gte-small";
+const extractor = await pipeline("feature-extraction", modelId);
+// see https://huggingface.co/thenlper/gte-small/blob/d8e2604cadbeeda029847d19759d219e0ce2e6d8/README.md?code=true#L2625
+export const MAX_SEQ_LEN = 512 as const;
 
 export async function findSimilarSentences(
 	query: string,