amruthpillai/reactive-resume · error · ORPCError

INTERNAL_SERVER_ERROR

INTERNAL_SERVER_ERROR

Error message

The AI response could not be parsed.

What it means

INTERNAL_SERVER_ERROR thrown by generateJson when the model's text response contains no parseable JSON object — specifically when there is no '{' ... '}' pair with end >= start after extracting a fenced ```json block or falling back to the raw text. It means the AI returned prose or empty content rather than JSON. Note the choice of INTERNAL_SERVER_ERROR (not BAD_REQUEST) because the input was fine; the model/parse pipeline failed.

Source

Thrown at packages/api/src/features/applications/ai.ts:61

	}
	return getModel({
		provider: provider.provider,
		model: provider.model,
		apiKey: provider.apiKey,
		...(provider.baseURL ? { baseURL: provider.baseURL } : {}),
	});
}

// generateText + tolerant JSON extraction + Zod validation. Mirrors the resume-analysis pattern
// (the SDK's generateObject isn't wired for every provider here, so we parse defensively).
async function generateJson<T>(model: Awaited<ReturnType<typeof resolveModel>>, prompt: string, schema: z.ZodType<T>) {
	const { text } = await generateText({ model, messages: [{ role: "user", content: prompt }] });
	const fenced = text.match(/```(?:json)?\s*([\s\S]*?)\s*```/);
	const candidate = fenced?.[1] ?? text;
	const start = candidate.indexOf("{");
	const end = candidate.lastIndexOf("}");
	if (start === -1 || end === -1 || end < start) {
		throw new ORPCError("INTERNAL_SERVER_ERROR", { message: "The AI response could not be parsed." });
	}
	return schema.parse(JSON.parse(candidate.slice(start, end + 1)));
}

async function generatePlainText(model: Awaited<ReturnType<typeof resolveModel>>, prompt: string) {
	const { text } = await generateText({ model, messages: [{ role: "user", content: prompt }] });
	return text.trim();
}

function isPrivateIPv4(address: string) {
	const parts = address.split(".").map((part) => Number(part));
	if (parts.length !== 4 || parts.some((part) => !Number.isInteger(part) || part < 0 || part > 255)) return true;
	const [a = 0, b = 0] = parts;
	return (
		a === 0 ||
		a === 10 ||
		a === 127 ||
		(a === 100 && b >= 64 && b <= 127) ||

View on GitHub (pinned to 3a5b12e2a4)

Solutions

1. Retry the request — transient model behavior often resolves on the next call.
2. Strengthen the prompt to demand pure JSON with no surrounding prose.
3. Switch to a more capable/instruction-following model for the applications feature.
4. If using a gateway, verify it forwards the model response verbatim and does not inject HTML error pages.
5. Handle this 500 in the UI with a friendly 'AI response was malformed, please try again' message.

Example fix

// before
prompt: 'Analyze this job.' // model returns prose

// after
prompt: 'Respond with ONLY raw JSON matching this schema, no markdown, no explanation:\n' + schemaExample

function extractJsonObject(text) {
  const fenced = text.match(/```(?:json)?\s*([\s\S]*?)\s*```/);
  const candidate = fenced?.[1] ?? text;
  const start = candidate.indexOf('{');
  const end = candidate.lastIndexOf('}');
  if (start === -1 || end === -1 || end < start) return null;
  try { return JSON.parse(candidate.slice(start, end + 1)); } catch { return null; }
}

function hasJsonObject(text) {
  return extractJsonObject(text) !== null;
}

try {
  await applications.analyzeJobPosting({ url });
} catch (e) {
  if (e.code === 'INTERNAL_SERVER_ERROR' && /could not be parsed/i.test(e.message)) {
    await backoffRetry(() => applications.analyzeJobPosting({ url }), { tries: 2 });
  } else throw e;
}

When it happens

Trigger: An applications AI feature calls generateJson; the model returns non-JSON text (apology, markdown without braces, empty string, or a refused response) so the brace-finding heuristic finds no candidate object.

Common situations: Model refuses due to content policy and returns prose; model is overloaded and returns empty; a provider gateway returns an error page as text; weak model ignoring the JSON instruction; prompt too long causing truncation before JSON.

Related errors

• The model returned too much text during the provider test.
• BAD_REQUEST
• An unknown error occurred while validating the merged resume
• BAD_GATEWAY
• BAD_REQUEST

AI-assisted analysis of amruthpillai/reactive-resume@3a5b12e2a4 (2026-08-12). Data as JSON: /api/errors/7eb24969261e0cca. Report an issue: GitHub.