fix(ai): Handle empty parts when streaming (#9262)

Author: hsubox76

.changeset/poor-cobras-dream.md

@@ -0,0 +1,5 @@
+---
+'@firebase/ai': patch
+---
+
+Updated SDK to handle empty parts when streaming.

packages/ai/src/googleai-mappers.ts

@@ -176,7 +176,7 @@ export function mapGenerateContentCandidates(
       // Throw early since developers may send a long video as input and only expect to pay
       // for inference on a small portion of the video.
       if (
-        candidate.content?.parts.some(
+        candidate.content?.parts?.some(
           part => (part as InlineDataPart)?.videoMetadata
         )
       ) {

packages/ai/src/methods/generate-content.test.ts

@@ -386,6 +386,22 @@ describe('generateContent()', () => {
       match.any
     );
   });
+  it('empty part', async () => {
+    const mockResponse = getMockResponse(
+      'vertexAI',
+      'unary-success-empty-part.json'
+    );
+    stub(request, 'makeRequest').resolves(mockResponse as Response);
+    const result = await generateContent(
+      fakeApiSettings,
+      'model',
+      fakeRequestParams
+    );
+    expect(result.response.text()).to.include(
+      'I can certainly help you with that!'
+    );
+    expect(result.response.inlineDataParts()?.length).to.equal(1);
+  });
   it('unknown enum - should ignore', async () => {
     const mockResponse = getMockResponse(
       'vertexAI',

packages/ai/src/requests/stream-reader.test.ts

@@ -194,6 +194,20 @@ describe('processStream', () => {
       expect(response.text()).to.equal('');
     }
   });
+  it('handles empty parts', async () => {
+    const fakeResponse = getMockResponseStreaming(
+      'googleAI',
+      'streaming-success-empty-parts.txt'
+    );
+
+    const result = processStream(fakeResponse as Response, fakeApiSettings);
+    for await (const response of result.stream) {
+      expect(response.candidates?.[0].content.parts.length).to.be.at.least(1);
+    }
+
+    const aggregatedResponse = await result.response;
+    expect(aggregatedResponse.candidates?.[0].content.parts.length).to.equal(6);
+  });
   it('unknown enum - should ignore', async () => {
     const fakeResponse = getMockResponseStreaming(
       'vertexAI',

packages/ai/src/requests/stream-reader.ts

@@ -100,6 +100,17 @@ async function* generateResponseSequence(
       enhancedResponse = createEnhancedContentResponse(value);
     }
 
+    const firstCandidate = enhancedResponse.candidates?.[0];
+    // Don't yield a response with no useful data for the developer.
+    if (
+      !firstCandidate?.content?.parts &&
+      !firstCandidate?.finishReason &&
+      !firstCandidate?.citationMetadata &&
+      !firstCandidate?.urlContextMetadata
+    ) {
+      continue;
+    }
+
     yield enhancedResponse;
   }
 }
@@ -211,37 +222,30 @@ export function aggregateResponses(
          * Candidates should always have content and parts, but this handles
          * possible malformed responses.
          */
-        if (candidate.content && candidate.content.parts) {
+        if (candidate.content) {
+          // Skip a candidate without parts.
+          if (!candidate.content.parts) {
+            continue;
+          }
           if (!aggregatedResponse.candidates[i].content) {
             aggregatedResponse.candidates[i].content = {
               role: candidate.content.role || 'user',
               parts: []
             };
           }
-          const newPart: Partial<Part> = {};
           for (const part of candidate.content.parts) {
-            if (part.text !== undefined) {
-              // The backend can send empty text parts. If these are sent back
-              // (e.g. in chat history), the backend will respond with an error.
-              // To prevent this, ignore empty text parts.
-              if (part.text === '') {
-                continue;
-              }
-              newPart.text = part.text;
-            }
-            if (part.functionCall) {
-              newPart.functionCall = part.functionCall;
+            const newPart: Part = { ...part };
+            // The backend can send empty text parts. If these are sent back
+            // (e.g. in chat history), the backend will respond with an error.
+            // To prevent this, ignore empty text parts.
+            if (part.text === '') {
+              continue;
             }
-            if (Object.keys(newPart).length === 0) {
-              throw new AIError(
-                AIErrorCode.INVALID_CONTENT,
-                'Part should have at least one property, but there are none. This is likely caused ' +
-                  'by a malformed response from the backend.'
+            if (Object.keys(newPart).length > 0) {
+              aggregatedResponse.candidates[i].content.parts.push(
+                newPart as Part
               );
             }
-            aggregatedResponse.candidates[i].content.parts.push(
-              newPart as Part
-            );
           }
         }
       }