Fix thinking models on Ollama

src/lib/ai/providers.test.ts

@@ -58,6 +58,14 @@ describe("OllamaProvider", () => {
     });
   });
 
+  it("never lets the model think, and puts the schema in the instructions for JSON", async () => {
+    const requests = stubFetch(() => ollamaChat('{"name":"Jane","years":9}'));
+    const provider = new OllamaProvider("http://ollama:11434", "qwen3.5:4b", 60_000);
+    await provider.complete({ ...ask, schema: Person });
+    expect(requests[0].body).toMatchObject({ think: false, format: { type: "object" } });
+    expect((requests[0].body as { messages: { content: string }[] }).messages[0].content).toMatch(/^Be brief\.\n\nReply with only a JSON object[\s\S]*"years"/);
+  });
+
   it("asks for JSON with the schema and validates the answer", async () => {
     const requests = stubFetch(() => ollamaChat('{"name":"Jane","years":9}'));
     const provider = new OllamaProvider("http://ollama:11434", "qwen2.5:3b", 60_000);

src/lib/ai/providers/ollama.ts

@@ -33,11 +33,18 @@ export class OllamaProvider extends BaseProvider {
   }
 
   protected async chat(call: ChatCall): Promise<ChatAnswer> {
+    const schema = call.json ? z.toJSONSchema(call.json.schema) : undefined;
+    // A model on local hardware can think past any token limit and never answer, so it doesn't think.
+    // One that doesn't think ignores `format`, so the schema goes into the instructions as well.
+    const system = schema
+      ? `${call.system}\n\nReply with only a JSON object, no other text, that matches this JSON schema:\n${JSON.stringify(schema)}`
+      : call.system;
     const response = await this.client(call.signal).chat({
       model: call.model,
-      messages: [{ role: "system", content: call.system }, ...call.messages],
+      messages: [{ role: "system", content: system }, ...call.messages],
       stream: false,
-      format: call.json ? z.toJSONSchema(call.json.schema) : undefined,
+      think: false,
+      format: schema,
       options: { temperature: call.temperature, num_predict: call.maxTokens, num_ctx: CONTEXT_TOKENS },
     });
     return { text: response.message.content, cutOff: response.done_reason === "length" };

src/lib/ai/tasks/extract-posting.ts

@@ -7,7 +7,8 @@ import { postingForPrompt, system, UNTRUSTED_POSTING, type Task } from "./prompt
 export const EXTRACT_POSTING_VERSION = "extract-posting@1";
 
 // What the model returns. The deadline is free text here and checked in
-// code, since models like to answer "15 Oct" rather than an ISO date.
+// code, since models like to answer "15 Oct" rather than an ISO date. Small
+// models leave out the keys of a requirement that don't apply, so those read as null.
 const answerSchema = z.object({
   title: z.string().nullable(),
   company: z.string().nullable(),
@@ -16,7 +17,13 @@ const answerSchema = z.object({
   mode: z.string().nullable(),
   salary: z.string().nullable(),
   deadline: z.string().nullable(),
-  requirements: z.array(requirementSchema),
+  requirements: z.array(
+    requirementSchema.extend({
+      skill: z.string().nullable().default(null),
+      years: z.number().nullable().default(null),
+      language: z.string().nullable().default(null),
+    }),
+  ),
   asksForReferences: z.boolean(),
   language: z.string().nullable(),
 });
@@ -25,6 +32,7 @@ const INSTRUCTIONS = `You read job postings and pull out their details as JSON.
 
 - Use only what the posting says; use null for anything it doesn't mention. Keep the posting's own wording for work mode and salary, e.g. "Hybrid · 3 days in office" or "€75,000–€90,000".
 - location: the city or region only, e.g. "Amsterdam".
+- company and role: the employer and the job title as the posting names them, often in its heading or first lines. null only when the posting names none.
 - title: "{role} · {company}", or null when either is missing.
 - deadline: the application deadline as YYYY-MM-DD. When the posting gives no year, it's the first such date after today.
 - requirements: one entry per thing the posting asks of the candidate, in its order, each shortened to the essentials (e.g. "5+ years of product design, ideally on complex B2B tools"). kind is:
@@ -38,16 +46,20 @@ const INSTRUCTIONS = `You read job postings and pull out their details as JSON.
 
 /** NEW-3: title, company, role, location, work mode, salary, deadline, typed requirements and language. */
 export const extractPosting: Task<{ posting: string; today?: Date }, Extracted> = async (provider, { posting, today = new Date() }, { signal }) => {
-  const answer = await provider.complete({
-    system: system(INSTRUCTIONS, `Today is ${localDate(today)}.`, UNTRUSTED_POSTING),
-    messages: [{ role: "user", content: postingForPrompt(posting) }],
-    schema: answerSchema,
-    schemaName: "extracted_posting",
-    // Reading, not writing: no creativity or long thought wanted.
-    creativity: "FAITHFUL",
-    effort: "low",
-    signal,
-  });
+  const ask = () =>
+    provider.complete({
+      system: system(INSTRUCTIONS, `Today is ${localDate(today)}.`, UNTRUSTED_POSTING),
+      messages: [{ role: "user", content: postingForPrompt(posting) }],
+      schema: answerSchema,
+      schemaName: "extracted_posting",
+      // Reading, not writing: no creativity or long thought wanted.
+      creativity: "FAITHFUL",
+      effort: "low",
+      signal,
+    });
+  let answer = await ask();
+  // A small model sometimes answers with every detail null while listing requirements; one more look usually finds them.
+  if (!answer.title && !answer.company && !answer.role) answer = await ask();
   const role = answer.role?.trim() || null;
   const company = answer.company?.trim() || null;
   return {

src/lib/ai/tasks/tasks.test.ts

@@ -76,6 +76,15 @@ describe("extractPosting", () => {
     expect(sent(provider).system).toContain("Today is 2026-09-28.");
   });
 
+  it("asks once more when the answer names no company, role or title, and takes requirements without their empty fields", async () => {
+    const empty = { ...extracted, title: null, company: null, role: null };
+    const replies = [empty, { ...extracted, requirements: [{ text: "Figma", kind: "skill", skill: "Figma" }] }];
+    const provider = new MockProvider({ reply: () => replies.shift() });
+    const { data } = await extractPosting(provider, { posting, today }, options);
+    expect(provider.calls).toHaveLength(2);
+    expect(data).toMatchObject({ company: "Ledgerly", requirements: [{ text: "Figma", years: null, language: null }] });
+  });
+
   it("only keeps whole dates", () => {
     expect(isoDate("2026-10-15")).toBe("2026-10-15");
     expect(isoDate("15 Oct 2026")).toBe("2026-10-15");