Merge pull request #94 from modelstudioai/feat/recommend-new

feat: remove intent-detect-v3 model

Author: gujieye

packages/commands/src/commands/advisor/recommend.ts

@@ -272,9 +272,7 @@ export default defineCommand({
       return result;
     });
 
-    const analyzeIntentPromise = analyzeIntent(ctx.client, userInput, {
-      intentDetectBaseUrl: settings.intentDetectBaseUrl,
-    }).then((result) => {
+    const analyzeIntentPromise = analyzeIntent(ctx.client, userInput).then((result) => {
       intentReady = true;
       if (!modelsReady) {
         spinner.update("Agent: Intent analyzed, loading model data...");

packages/commands/tests/e2e/advisor-recommend.e2e.test.ts

@@ -96,9 +96,9 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: advisor recommend (DashScope)", ()
     expect(data.result?.recommendations?.[0]?.highlights?.length).toBeGreaterThan(0);
   }, 120_000);
 
-  // ---- Model preference: positive cases ----
+  // ---- Mode coverage: all 4 modes ----
 
-  test("scoped preference — intent contains modelPreference.mode=scoped when family is specified", async () => {
+  test("mode: scoped — family-scoped query sets mode=scoped with targets", async () => {
     const { stdout, stderr, exitCode } = await runCommandE2e(ADVISOR_ROUTES, [
       "advisor",
       "recommend",
@@ -112,39 +112,62 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: advisor recommend (DashScope)", ()
     const data = parseStdoutJson<{
       intent?: { modelPreference?: { mode?: string; targets?: string[] } };
     }>(stdout);
-    // Model preference detection depends on LLM interpretation
-    // Accept either "scoped" or "unconstrained" as valid
-    const mode = data.intent?.modelPreference?.mode;
-    expect(mode === "scoped" || mode === "unconstrained" || mode === undefined).toBe(true);
+    const pref = data.intent?.modelPreference;
+    expect(pref?.mode).toBe("scoped");
+    expect(pref?.targets?.length).toBeGreaterThan(0);
+    // Should contain "deepseek" (case-insensitive substring)
+    expect(pref?.targets?.some((t) => t.toLowerCase().includes("deepseek"))).toBe(true);
   }, 60_000);
 
-  test("comparison preference — intent contains modelPreference.mode=comparison when comparing models", async () => {
+  test("mode: comparison — comparing two models sets mode=comparison with both targets", async () => {
     const { stdout, stderr, exitCode } = await runCommandE2e(ADVISOR_ROUTES, [
       "advisor",
       "recommend",
       "--dry-run",
       "--message",
-      "Which is better for code generation, qwen-max or deepseek-v3?",
+      "Compare qwen-max and deepseek-v3 for legal contract review, high precision required",
       "--output",
       "json",
     ]);
     expect(exitCode, stderr).toBe(0);
     const data = parseStdoutJson<{
       intent?: { modelPreference?: { mode?: string; targets?: string[] } };
     }>(stdout);
-    // Model preference detection depends on LLM interpretation
-    // Accept either "comparison" or "unconstrained" as valid
-    const mode = data.intent?.modelPreference?.mode;
-    expect(mode === "comparison" || mode === "unconstrained" || mode === undefined).toBe(true);
+    const pref = data.intent?.modelPreference;
+    expect(pref?.mode).toBe("comparison");
+    expect(pref?.targets?.length).toBeGreaterThanOrEqual(2);
+    const targetsLower = pref?.targets?.map((t) => t.toLowerCase()) ?? [];
+    expect(targetsLower.some((t) => t.includes("qwen"))).toBe(true);
+    expect(targetsLower.some((t) => t.includes("deepseek"))).toBe(true);
+  }, 60_000);
+
+  test("mode: alternative — reference model query sets mode=alternative with target", async () => {
+    const { stdout, stderr, exitCode } = await runCommandE2e(ADVISOR_ROUTES, [
+      "advisor",
+      "recommend",
+      "--dry-run",
+      "--message",
+      "Something like qwen-max but cheaper, for text summarization",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{
+      intent?: { modelPreference?: { mode?: string; targets?: string[] } };
+    }>(stdout);
+    const pref = data.intent?.modelPreference;
+    expect(pref?.mode).toBe("alternative");
+    expect(pref?.targets?.length).toBeGreaterThan(0);
+    expect(pref?.targets?.some((t) => t.toLowerCase().includes("qwen"))).toBe(true);
   }, 60_000);
 
-  test("excludes preference — intent detects modelPreference when excluding models", async () => {
+  test("mode: excludes — excluding a family populates excludes array", async () => {
     const { stdout, stderr, exitCode } = await runCommandE2e(ADVISOR_ROUTES, [
       "advisor",
       "recommend",
       "--dry-run",
       "--message",
-      "Not qwen, recommend a model suitable for text generation",
+      "Recommend a model for text generation, but not qwen",
       "--output",
       "json",
     ]);
@@ -157,18 +180,12 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: advisor recommend (DashScope)", ()
         };
       };
     }>(stdout);
-    // Model preference detection depends on LLM interpretation
-    // If excludes is detected, verify it contains qwen; otherwise accept as valid
     const pref = data.intent?.modelPreference;
-    if (pref?.excludes && pref.excludes.length > 0) {
-      expect(pref.excludes.some((e) => e.toLowerCase().includes("qwen"))).toBe(true);
-    }
-    // Test passes if exit code is 0, regardless of whether excludes was detected
+    expect(pref?.excludes?.length).toBeGreaterThan(0);
+    expect(pref?.excludes?.some((e) => e.toLowerCase().includes("qwen"))).toBe(true);
   }, 60_000);
 
-  // ---- Model preference: negative cases ----
-
-  test("no preference — intent has no modelPreference or mode=unconstrained for generic queries", async () => {
+  test("mode: unconstrained — generic query has no modelPreference", async () => {
     const { stdout, stderr, exitCode } = await runCommandE2e(ADVISOR_ROUTES, [
       "advisor",
       "recommend",

packages/core/src/advisor/constants/index.ts

@@ -1,8 +1,5 @@
 export { DEFAULT_INTENT } from "./defaults.ts";
 export {
-  INTENT_DETECT_MODEL,
-  INTENT_DETECT_TOOL,
-  buildIntentDetectSystemPrompt,
   INTENT_EXTRACTION_MODEL,
   INTENT_SYSTEM_PROMPT,
   JSON_RETRY_HINT,

packages/core/src/advisor/constants/prompts.ts

@@ -1,14 +1,9 @@
 export const RANKING_MODEL = "qwen-flash";
 
 /**
- * Dedicated intent-detection model. Sub-100ms latency, designed for fast
- * classification + tool routing. Provides mode/targets/excludes/complexity.
- */
-export const INTENT_DETECT_MODEL = "tongyi-intent-detect-v3";
-
-/**
- * Rich field extraction model. Runs in parallel with detect-v3 to extract
- * taskSummary, modalities, budget, qualityPreference, etc.
+ * Intent extraction model. Runs a single call to extract taskSummary,
+ * modalities, budget, qualityPreference, modelPreference, and all other
+ * structured fields from the user's input.
  */
 export const INTENT_EXTRACTION_MODEL = "qwen3.6-flash";
 
@@ -193,74 +188,3 @@ The intent's modelPreference.targets is the reference model.
 
 ## Output Format
 {"type":"single","recommendations":[{"model":"model ID","reason":"alternative analysis","highlights":["differentiators"]}]}`;
-
-/**
- * Tool definition for `tongyi-intent-detect-v3`. Serialized to a JSON string
- * and embedded in the system prompt (NOT passed via the request body `tools`
- * field — this model doesn't use OpenAI function-calling; it has its own
- * `<tags>` / `
-</think>
-
-` output format driven by the system prompt).
- */
-export const INTENT_DETECT_TOOL = {
-  name: "classify_intent",
-  description:
-    "Classify the user's model recommendation intent. Extract the mode and any model/family names mentioned.",
-  parameters: {
-    type: "object",
-    properties: {
-      mode: {
-        type: "string",
-        enum: ["unconstrained", "scoped", "comparison", "alternative"],
-        description: "The detected intent mode.",
-      },
-      targets: {
-        type: "array",
-        items: { type: "string" },
-        description:
-          "Model or family names the user mentioned or wants to evaluate. Empty for unconstrained.",
-      },
-      excludes: {
-        type: "array",
-        items: { type: "string" },
-        description:
-          "Model or family names the user explicitly wants to exclude. Empty if none mentioned.",
-      },
-      complexity: {
-        type: "string",
-        enum: ["single", "pipeline"],
-        description:
-          "Whether the task needs a single model or a multi-step pipeline. Default to single unless the user clearly describes chained steps.",
-      },
-    },
-    required: ["mode"],
-  },
-} as const;
-
-/**
- * Build the system prompt for `tongyi-intent-detect-v3` following the official
- * template from the model's documentation. The tools JSON is embedded in the
- * prompt text — the model reads it from the system message, not from a separate
- * `tools` request field.
- *
- * Official template:
- *   "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
- *    You may call one or more tools to assist with the user query.
- *    The tools you can use are as follows:
- *    {tools_string}
- *    Response in INTENT_MODE."
- *
- * `INTENT_MODE` tells the model to emit `<tags>label</tags>` + `
-</think>
-
-`
- * output. The tag carries the mode classification; the tool_call carries
- * structured targets/excludes/complexity extracted from the user's prompt.
- */
-export function buildIntentDetectSystemPrompt(): string {
-  const toolsString = JSON.stringify([INTENT_DETECT_TOOL], null, 2);
-  return `You are Qwen, created by Alibaba Cloud. You are a helpful assistant. You may call one or more tools to assist with the user query. The tools you can use are as follows:
-${toolsString}
-Response in INTENT_MODE.`;
-}