fix(server): rerank scores calc (#13016)

fix AI-257

packages/backend/server/src/plugins/copilot/providers/gemini/gemini.ts

@@ -5,6 +5,7 @@ import type {
 import type { GoogleVertexProvider } from '@ai-sdk/google-vertex';
 import {
   AISDKError,
+  embedMany,
   generateObject,
   generateText,
   JSONParseError,
@@ -20,6 +21,7 @@ import {
 import { CopilotProvider } from '../provider';
 import type {
   CopilotChatOptions,
+  CopilotEmbeddingOptions,
   CopilotImageOptions,
   CopilotProviderModel,
   ModelConditions,
@@ -211,6 +213,40 @@ export abstract class GeminiProvider<T> extends CopilotProvider<T> {
     }
   }
 
+  override async embedding(
+    cond: ModelConditions,
+    messages: string | string[],
+    options: CopilotEmbeddingOptions = { dimensions: DEFAULT_DIMENSIONS }
+  ): Promise<number[][]> {
+    messages = Array.isArray(messages) ? messages : [messages];
+    const fullCond = { ...cond, outputType: ModelOutputType.Embedding };
+    await this.checkParams({ embeddings: messages, cond: fullCond, options });
+    const model = this.selectModel(fullCond);
+
+    try {
+      metrics.ai
+        .counter('generate_embedding_calls')
+        .add(1, { model: model.id });
+
+      const modelInstance = this.instance.textEmbeddingModel(model.id, {
+        outputDimensionality: options.dimensions || DEFAULT_DIMENSIONS,
+        taskType: 'RETRIEVAL_DOCUMENT',
+      });
+
+      const { embeddings } = await embedMany({
+        model: modelInstance,
+        values: messages,
+      });
+
+      return embeddings.filter(v => v && Array.isArray(v));
+    } catch (e: any) {
+      metrics.ai
+        .counter('generate_embedding_errors')
+        .add(1, { model: model.id });
+      throw this.handleError(e);
+    }
+  }
+
   private async getFullStream(
     model: CopilotProviderModel,
     messages: PromptMessage[],

packages/backend/server/src/plugins/copilot/providers/gemini/generative.ts

@@ -71,8 +71,8 @@ export class GeminiGenerativeProvider extends GeminiProvider<GeminiGenerativeCon
       ],
     },
     {
-      name: 'Text Embedding 004',
-      id: 'text-embedding-004',
+      name: 'Text Embedding 005',
+      id: 'text-embedding-005',
       capabilities: [
         {
           input: [ModelInputType.Text],
@@ -80,6 +80,18 @@ export class GeminiGenerativeProvider extends GeminiProvider<GeminiGenerativeCon
         },
       ],
     },
+    // not exists yet
+    // {
+    //   name: 'Gemini Embedding',
+    //   id: 'gemini-embedding-001',
+    //   capabilities: [
+    //     {
+    //       input: [ModelInputType.Text],
+    //       output: [ModelOutputType.Embedding],
+    //       defaultForOutputType: true,
+    //     },
+    //   ],
+    // },
   ];
 
   protected instance!: GoogleGenerativeAIProvider;

packages/backend/server/src/plugins/copilot/providers/gemini/vertex.ts

@@ -49,6 +49,17 @@ export class GeminiVertexProvider extends GeminiProvider<GeminiVertexConfig> {
         },
       ],
     },
+    {
+      name: 'Gemini Embedding',
+      id: 'gemini-embedding-001',
+      capabilities: [
+        {
+          input: [ModelInputType.Text],
+          output: [ModelOutputType.Embedding],
+          defaultForOutputType: true,
+        },
+      ],
+    },
   ];
 
   protected instance!: GoogleVertexProvider;

packages/backend/server/src/plugins/copilot/providers/openai.ts

@@ -450,7 +450,8 @@ export class OpenAIProvider extends CopilotProvider<OpenAIConfig> {
     const fullCond = { ...cond, outputType: ModelOutputType.Text };
     await this.checkParams({ messages: [], cond: fullCond, options });
     const model = this.selectModel(fullCond);
-    const instance = this.#instance.responses(model.id);
+    // get the log probability of "yes"/"no"
+    const instance = this.#instance(model.id, { logprobs: 16 });
 
     const scores = await Promise.all(
       chunkMessages.map(async messages => {
@@ -461,29 +462,37 @@ export class OpenAIProvider extends CopilotProvider<OpenAIConfig> {
           system,
           messages: msgs,
           temperature: 0,
-          maxTokens: 1,
+          maxTokens: 16,
           providerOptions: {
             openai: {
               ...this.getOpenAIOptions(options, model.id),
-              // get the log probability of "yes"/"no"
-              logprobs: 2,
             },
           },
-          maxSteps: 1,
           abortSignal: options.signal,
         });
 
-        const top = (logprobs?.[0]?.topLogprobs ?? []).reduce(
-          (acc, item) => {
-            acc[item.token] = item.logprob;
-            return acc;
-          },
-          {} as Record<string, number>
+        const topMap: Record<string, number> = (
+          logprobs?.[0]?.topLogprobs ?? []
+        ).reduce<Record<string, number>>(
+          (acc, { token, logprob }) => ({ ...acc, [token]: logprob }),
+          {}
         );
 
-        // OpenAI often includes a leading space, so try matching both ' yes' and 'yes'
-        const logYes = top[' yes'] ?? top['yes'] ?? Number.NEGATIVE_INFINITY;
-        const logNo = top[' no'] ?? top['no'] ?? Number.NEGATIVE_INFINITY;
+        const findLogProb = (token: string): number => {
+          // OpenAI often includes a leading space, so try matching '.yes', '_yes', ' yes' and 'yes'
+          return [`.${token}`, `_${token}`, ` ${token}`, token]
+            .flatMap(v => [v, v.toLowerCase(), v.toUpperCase()])
+            .reduce<number>(
+              (best, key) =>
+                (topMap[key] ?? Number.NEGATIVE_INFINITY) > best
+                  ? topMap[key]
+                  : best,
+              Number.NEGATIVE_INFINITY
+            );
+        };
+
+        const logYes = findLogProb('Yes');
+        const logNo = findLogProb('No');
 
         const pYes = Math.exp(logYes);
         const pNo = Math.exp(logNo);

packages/backend/server/src/plugins/copilot/providers/utils.ts

@@ -501,6 +501,12 @@ export class TextStreamParser {
           case 'doc_semantic_search': {
             if (Array.isArray(chunk.result)) {
               result += `\nFound ${chunk.result.length} document${chunk.result.length !== 1 ? 's' : ''} related to “${chunk.args.query}”.\n`;
+            } else if (typeof chunk.result === 'string') {
+              result += `\n${chunk.result}\n`;
+            } else {
+              this.logger.warn(
+                `Unexpected result type for doc_semantic_search: ${chunk.result?.message || 'Unknown error'}`
+              );
             }
             break;
           }