Merge branch 'feature/distillation-pipeline' into fine-tuning/student-explainability

# Conflicts:
#	.gitignore
#	codebenders-dashboard/app/api/students/[guid]/sis-link/route.ts

Author: William-Hill

.gitignore

@@ -186,3 +186,6 @@ operations/verify_institution_id.py
 
 # Test upload fixtures (generated — do not commit)
 data/test_uploads/
+
+# Training pipeline artifacts
+training_data/

codebenders-dashboard/app/api/courses/explain-pairing/route.ts

@@ -1,10 +1,7 @@
 import { type NextRequest, NextResponse } from "next/server"
 import { getPool } from "@/lib/db"
 import { canAccess, type Role } from "@/lib/roles"
-import { generateText } from "ai"
-import { createOpenAI } from "@ai-sdk/openai"
-
-const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY || "" })
+import { generateExplanation } from "@/lib/model-client"
 
 const DELIVERY_LABELS: Record<string, string> = {
   F: "Face-to-Face",
@@ -18,7 +15,7 @@ export async function POST(request: NextRequest) {
     return NextResponse.json({ error: "Forbidden" }, { status: 403 })
   }
 
-  if (!process.env.OPENAI_API_KEY) {
+  if (process.env.MODEL_BACKEND !== "ollama" && !process.env.OPENAI_API_KEY) {
     return NextResponse.json({ error: "OpenAI API key not configured" }, { status: 500 })
   }
 
@@ -189,11 +186,7 @@ Write a concise analysis (3-4 sentences) that:
 
 Be practical and data-driven. Do not speculate beyond what the numbers show.`
 
-    const result = await generateText({
-      model: openai("gpt-4o-mini"),
-      prompt: llmPrompt,
-      maxOutputTokens: 320,
-    })
+    const result = { text: await generateExplanation(llmPrompt, 320) }
 
     return NextResponse.json({ stats, explanation: result.text })
   } catch (error) {

codebenders-dashboard/app/api/query-summary/route.ts

@@ -1,17 +1,14 @@
 import { type NextRequest, NextResponse } from "next/server"
 import { canAccess, type Role } from "@/lib/roles"
-import { generateText } from "ai"
-import { createOpenAI } from "@ai-sdk/openai"
-
-const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY || "" })
+import { generateSummary } from "@/lib/model-client"
 
 export async function POST(request: NextRequest) {
   const role = request.headers.get("x-user-role") as Role | null
   if (!role || !canAccess("/api/query-summary", role)) {
     return NextResponse.json({ error: "Forbidden" }, { status: 403 })
   }
 
-  if (!process.env.OPENAI_API_KEY) {
+  if (process.env.MODEL_BACKEND !== "ollama" && !process.env.OPENAI_API_KEY) {
     return NextResponse.json({ error: "OpenAI API key not configured" }, { status: 500 })
   }
 
@@ -47,12 +44,8 @@ ${JSON.stringify(sampleRows, null, 2)}
 Write a 2-3 sentence plain-English summary of what these results show. Be specific about the numbers. Do not speculate beyond the data. Address the advisor directly.`
 
   try {
-    const result = await generateText({
-      model: openai("gpt-4o-mini"),
-      prompt: llmPrompt,
-      maxOutputTokens: 200,
-    })
-    return NextResponse.json({ summary: result.text })
+    const summary = await generateSummary(llmPrompt, 200)
+    return NextResponse.json({ summary })
   } catch (error) {
     console.error("[query-summary] Error:", error)
     return NextResponse.json(

codebenders-dashboard/lib/model-client.ts

@@ -0,0 +1,81 @@
+/**
+ * Model client adapter — routes inference to Ollama (fine-tuned) or
+ * OpenAI (fallback) based on MODEL_BACKEND env var.
+ */
+
+import { generateText } from "ai"
+import { createOpenAI } from "@ai-sdk/openai"
+
+const MODEL_BACKEND = process.env.MODEL_BACKEND || "openai"
+const SCHOOL_CODE = process.env.SCHOOL_CODE || "bishop-state"
+const OLLAMA_BASE_URL = process.env.OLLAMA_BASE_URL || "http://localhost:11434"
+const MODEL_SIZE = process.env.MODEL_SIZE || "9b"
+
+let _openai: ReturnType<typeof createOpenAI> | null = null
+
+function getOpenAI() {
+  if (!_openai) {
+    _openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY || "" })
+  }
+  return _openai
+}
+
+async function callOllama(model: string, prompt: string, maxTokens: number): Promise<string> {
+  const response = await fetch(`${OLLAMA_BASE_URL}/api/generate`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      model,
+      prompt,
+      stream: false,
+      options: {
+        temperature: 0.3,
+        num_predict: maxTokens,
+      },
+    }),
+  })
+
+  if (!response.ok) {
+    throw new Error(`Ollama error: ${response.status} ${response.statusText}`)
+  }
+
+  const data = await response.json()
+  return data.response
+}
+
+async function generate(
+  task: "explainer" | "summarizer",
+  prompt: string,
+  maxTokens: number,
+): Promise<string> {
+  if (MODEL_BACKEND === "ollama") {
+    const model = `${SCHOOL_CODE}-${task}:${MODEL_SIZE}`
+    return callOllama(model, prompt, maxTokens)
+  }
+  const result = await generateText({
+    model: getOpenAI()("gpt-4o-mini"),
+    prompt,
+    maxOutputTokens: maxTokens,
+  })
+  return result.text
+}
+
+/**
+ * Generate a course pairing explanation.
+ */
+export async function generateExplanation(
+  prompt: string,
+  maxTokens: number = 320,
+): Promise<string> {
+  return generate("explainer", prompt, maxTokens)
+}
+
+/**
+ * Generate a query result summary.
+ */
+export async function generateSummary(
+  prompt: string,
+  maxTokens: number = 200,
+): Promise<string> {
+  return generate("summarizer", prompt, maxTokens)
+}