diff --git a/knowledge-fs/apps/api/src/answer-generation-options.test.ts b/knowledge-fs/apps/api/src/answer-generation-options.test.ts
index 0732e2a8414..7d8d1b7d79e 100644
--- a/knowledge-fs/apps/api/src/answer-generation-options.test.ts
+++ b/knowledge-fs/apps/api/src/answer-generation-options.test.ts
@@ -113,6 +113,42 @@ describe("createApiAnswerGenerationOptions", () => {
     expect(indexSource).not.toContain("model: answerGenerationOptions.model");
   });
 
+  it("keeps interactive retrieval evidence-only and reserves LLM synthesis for Research", async () => {
+    const { readFile } = await import("node:fs/promises");
+    const indexSource = await readFile(new URL("./index.ts", import.meta.url), "utf8");
+    const citationOptionsStart = indexSource.indexOf("const multimodalCitationOptions =");
+    const researchMultimodalOptionsStart = indexSource.indexOf(
+      "const researchAnswerMultimodalOptions =",
+    );
+    const retrievalAssemblyStart = indexSource.indexOf(
+      "const retrievalEvidenceQueryGenerator = retriever",
+    );
+    const researchAnswerAssemblyStart = indexSource.indexOf(
+      "const profileLlmAnswerQueryGenerator = retriever",
+    );
+    const retrievalAssembly = indexSource.slice(
+      retrievalAssemblyStart,
+      researchAnswerAssemblyStart,
+    );
+    const citationOptionsAssembly = indexSource.slice(
+      citationOptionsStart,
+      researchMultimodalOptionsStart,
+    );
+
+    expect(citationOptionsStart).toBeGreaterThanOrEqual(0);
+    expect(researchMultimodalOptionsStart).toBeGreaterThan(citationOptionsStart);
+    expect(retrievalAssemblyStart).toBeGreaterThanOrEqual(0);
+    expect(researchAnswerAssemblyStart).toBeGreaterThan(retrievalAssemblyStart);
+    expect(citationOptionsAssembly).toContain("multimodalCandidateResolver");
+    expect(citationOptionsAssembly).not.toContain("multimodalAnswerOptions");
+    expect(retrievalAssembly).toContain("createHybridQueryGenerator");
+    expect(retrievalAssembly).toContain("...multimodalCitationOptions");
+    expect(retrievalAssembly).not.toContain("multimodalAnswerOptions");
+    expect(indexSource).toContain("generator: researchAnswerQueryGenerator");
+    expect(indexSource).toContain("{ queryGenerator: retrievalEvidenceQueryGenerator }");
+    expect(indexSource).not.toContain("{ queryGenerator: researchAnswerQueryGenerator }");
+  });
+
   it("honors explicit output-token overrides", () => {
     const options = createApiAnswerGenerationOptions({
       KNOWLEDGE_ANSWER_MAX_OUTPUT_TOKENS: "2048",
diff --git a/knowledge-fs/apps/api/src/index.ts b/knowledge-fs/apps/api/src/index.ts
index 0483c57d1e6..6ed1e37bf92 100644
--- a/knowledge-fs/apps/api/src/index.ts
+++ b/knowledge-fs/apps/api/src/index.ts
@@ -558,8 +558,8 @@ const retrievalTestExecutor = retriever
       retriever,
     })
   : undefined;
-// Resolves multimodal citations (manifest item id, asset route, page/bbox) and is shared by both
-// query generators. Available in database-repository mode, which is also when the retriever exists.
+// Resolves multimodal citations (manifest item id, asset route, page/bbox) without invoking a
+// model. Available in database-repository mode, which is also when the retriever exists.
 const multimodalCandidateResolver =
   repositoryOptions.documentAssets && repositoryOptions.parseArtifacts
     ? createDocumentMultimodalCandidateResolver({
@@ -570,21 +570,26 @@ const multimodalCandidateResolver =
         parseArtifacts: repositoryOptions.parseArtifacts,
       })
     : undefined;
-const multimodalGeneratorOptions = {
-  ...multimodalAnswerOptions,
+const multimodalCitationOptions = {
   ...(multimodalCandidateResolver ? { multimodalCandidateResolver } : {}),
 };
-const extractiveQueryGenerator = retriever
+const researchAnswerMultimodalOptions = {
+  ...multimodalAnswerOptions,
+  ...multimodalCitationOptions,
+};
+// Interactive query-stream is a retrieval surface: it returns bounded evidence and citations,
+// never an LLM/VLM-synthesized answer. Answer synthesis remains a separate Research capability.
+const retrievalEvidenceQueryGenerator = retriever
   ? createHybridQueryGenerator({
       limit: 5,
       maxAnswerChars: 2_000,
       retriever,
       topK: 10,
       ...embeddingGeneratorOptions,
-      ...multimodalGeneratorOptions,
+      ...multimodalCitationOptions,
     })
   : undefined;
-const llmAnswerQueryGenerator = retriever
+const profileLlmAnswerQueryGenerator = retriever
   ? createLlmAnswerQueryGenerator({
       limit: 5,
       maxAnswerChars: 2_000,
@@ -594,14 +599,14 @@ const llmAnswerQueryGenerator = retriever
       temperature: 0,
       topK: 10,
       ...embeddingGeneratorOptions,
-      ...multimodalGeneratorOptions,
+      ...researchAnswerMultimodalOptions,
     })
   : undefined;
-const queryGenerator =
-  extractiveQueryGenerator && llmAnswerQueryGenerator
+const researchAnswerQueryGenerator =
+  retrievalEvidenceQueryGenerator && profileLlmAnswerQueryGenerator
     ? createProfileAwareQueryGenerator({
-        extractiveGenerator: extractiveQueryGenerator,
-        profileLlmGenerator: llmAnswerQueryGenerator,
+        extractiveGenerator: retrievalEvidenceQueryGenerator,
+        profileLlmGenerator: profileLlmAnswerQueryGenerator,
       })
     : undefined;
 const researchProjectionSnapshotResolver = repositoryOptions.projectionSetPublications
@@ -646,7 +651,7 @@ const researchTaskRuntime =
   databaseRepositories.researchTaskPartialResults &&
   databaseRepositories.researchTaskProgressEvents &&
   repositoryOptions.knowledgeSpaceAccess &&
-  queryGenerator
+  researchAnswerQueryGenerator
     ? createApiResearchTaskRuntime({
         access: repositoryOptions.knowledgeSpaceAccess,
         adapter,
@@ -654,7 +659,7 @@ const researchTaskRuntime =
           ? { capabilityGrants: databaseRepositories.capabilityGrantProvenance }
           : {}),
         ...(deletionLifecycleFence ? { deletionFence: deletionLifecycleFence } : {}),
-        generator: queryGenerator,
+        generator: researchAnswerQueryGenerator,
         manifests: knowledgeSpaceManifests,
         metrics: operationalMetrics.durableTasks,
         partials: databaseRepositories.researchTaskPartialResults,
@@ -768,7 +773,7 @@ const app = createKnowledgeGateway({
         },
       }
     : {}),
-  ...(queryGenerator ? { queryGenerator } : {}),
+  ...(retrievalEvidenceQueryGenerator ? { queryGenerator: retrievalEvidenceQueryGenerator } : {}),
   ...(retrievalTestExecutor ? { retrievalTestExecutor } : {}),
   ...(publishedGraph ? { publishedGraph } : {}),
   ...(researchTaskRuntime