Implement dual-model filing pipeline with Ollama extraction

lib/server/ai.ts

@@ -1,7 +1,12 @@
+import { createOpenAI } from '@ai-sdk/openai';
 import { generateText } from 'ai';
 import { createZhipu } from 'zhipu-ai-provider';
 
+type AiWorkload = 'report' | 'extraction';
+type AiProvider = 'zhipu' | 'ollama';
+
 type AiConfig = {
+  provider: AiProvider;
   apiKey?: string;
   baseUrl: string;
   model: string;
@@ -27,11 +32,15 @@ type AiGenerateOutput = {
 };
 
 type RunAiAnalysisOptions = GetAiConfigOptions & {
+  workload?: AiWorkload;
   createModel?: (config: AiConfig) => unknown;
   generate?: (input: AiGenerateInput) => Promise<AiGenerateOutput>;
 };
 
 const CODING_API_BASE_URL = 'https://api.z.ai/api/coding/paas/v4';
+const OLLAMA_BASE_URL = 'http://127.0.0.1:11434';
+const OLLAMA_MODEL = 'qwen3:8b';
+const OLLAMA_API_KEY = 'ollama';
 
 let warnedIgnoredZhipuBaseUrl = false;
 
@@ -74,20 +83,47 @@ function fallbackResponse(prompt: string) {
   const clipped = prompt.split('\n').slice(0, 6).join(' ').slice(0, 260);
 
   return [
-    'AI SDK fallback mode is active (Zhipu configuration is missing).',
+    'AI SDK fallback mode is active (live model configuration is missing or unavailable).',
     'Thesis: Portfolio remains analyzable with local heuristics until live model access is configured.',
     'Risk scan: Concentration and filing sentiment should be monitored after each sync cycle.',
     `Context digest: ${clipped}`
   ].join('\n\n');
 }
 
+function toOpenAiCompatibleBaseUrl(baseUrl: string) {
+  const normalized = baseUrl.endsWith('/')
+    ? baseUrl.slice(0, -1)
+    : baseUrl;
+
+  return normalized.endsWith('/v1')
+    ? normalized
+    : `${normalized}/v1`;
+}
+
+function asErrorMessage(error: unknown) {
+  if (error instanceof Error && error.message) {
+    return error.message;
+  }
+
+  return String(error);
+}
+
 function defaultCreateModel(config: AiConfig) {
-  const zhipu = createZhipu({
-    apiKey: config.apiKey,
-    baseURL: config.baseUrl
+  if (config.provider === 'zhipu') {
+    const zhipu = createZhipu({
+      apiKey: config.apiKey,
+      baseURL: config.baseUrl
+    });
+
+    return zhipu(config.model);
+  }
+
+  const openai = createOpenAI({
+    apiKey: config.apiKey ?? OLLAMA_API_KEY,
+    baseURL: toOpenAiCompatibleBaseUrl(config.baseUrl)
   });
 
-  return zhipu(config.model);
+  return openai.chat(config.model);
 }
 
 async function defaultGenerate(input: AiGenerateInput): Promise<AiGenerateOutput> {
@@ -102,10 +138,15 @@ async function defaultGenerate(input: AiGenerateInput): Promise<AiGenerateOutput
 }
 
 export function getAiConfig(options?: GetAiConfigOptions) {
+  return getReportAiConfig(options);
+}
+
+export function getReportAiConfig(options?: GetAiConfigOptions) {
   const env = options?.env ?? process.env;
   warnIgnoredZhipuBaseUrl(env, options?.warn ?? console.warn);
 
   return {
+    provider: 'zhipu',
     apiKey: envValue('ZHIPU_API_KEY', env),
     baseUrl: CODING_API_BASE_URL,
     model: envValue('ZHIPU_MODEL', env) ?? 'glm-4.7-flashx',
@@ -113,15 +154,30 @@ export function getAiConfig(options?: GetAiConfigOptions) {
   } satisfies AiConfig;
 }
 
+export function getExtractionAiConfig(options?: GetAiConfigOptions) {
+  const env = options?.env ?? process.env;
+
+  return {
+    provider: 'ollama',
+    apiKey: envValue('OLLAMA_API_KEY', env) ?? OLLAMA_API_KEY,
+    baseUrl: envValue('OLLAMA_BASE_URL', env) ?? OLLAMA_BASE_URL,
+    model: envValue('OLLAMA_MODEL', env) ?? OLLAMA_MODEL,
+    temperature: 0
+  } satisfies AiConfig;
+}
+
 export function isAiConfigured(options?: GetAiConfigOptions) {
-  const config = getAiConfig(options);
+  const config = getReportAiConfig(options);
   return Boolean(config.apiKey);
 }
 
 export async function runAiAnalysis(prompt: string, systemPrompt?: string, options?: RunAiAnalysisOptions) {
-  const config = getAiConfig(options);
+  const workload = options?.workload ?? 'report';
+  const config = workload === 'extraction'
+    ? getExtractionAiConfig(options)
+    : getReportAiConfig(options);
 
-  if (!config.apiKey) {
+  if (workload === 'report' && !config.apiKey) {
     return {
       provider: 'local-fallback',
       model: config.model,
@@ -131,25 +187,49 @@ export async function runAiAnalysis(prompt: string, systemPrompt?: string, optio
 
   const createModel = options?.createModel ?? defaultCreateModel;
   const generate = options?.generate ?? defaultGenerate;
-  const model = createModel(config);
+  const warn = options?.warn ?? console.warn;
 
-  const result = await generate({
-    model,
-    system: systemPrompt,
-    prompt,
-    temperature: config.temperature
-  });
+  try {
+    const model = createModel(config);
 
-  const text = result.text.trim();
-  if (!text) {
-    throw new Error('AI SDK returned an empty response');
+    const result = await generate({
+      model,
+      system: systemPrompt,
+      prompt,
+      temperature: config.temperature
+    });
+
+    const text = result.text.trim();
+    if (!text) {
+      if (workload === 'extraction') {
+        return {
+          provider: 'local-fallback',
+          model: config.model,
+          text: fallbackResponse(prompt)
+        };
+      }
+
+      throw new Error('AI SDK returned an empty response');
+    }
+
+    return {
+      provider: config.provider,
+      model: config.model,
+      text
+    };
+  } catch (error) {
+    if (workload === 'extraction') {
+      warn(`[AI SDK] Extraction fallback activated: ${asErrorMessage(error)}`);
+
+      return {
+        provider: 'local-fallback',
+        model: config.model,
+        text: fallbackResponse(prompt)
+      };
+    }
+
+    throw error;
   }
-
-  return {
-    provider: 'zhipu',
-    model: config.model,
-    text
-  };
 }
 
 export function __resetAiWarningsForTests() {