Repairing Truncated LLM JSON

Quick answer: When an LLM hits its output-token limit mid-response, the JSON is cut off — strings and brackets are left open, and JSON.parse() throws Unexpected end of JSON input. To recover, either raise the token limit and re-run, or apply a repair pass that closes the dangling structures. The AI JSON Cleaner attempts to close open brackets and strings so you can salvage the partial result.

Truncation has a distinct signature, different from a syntax typo:

• The error is Unexpected end of JSON input (ran out of data), not Unexpected token (wrong character).
• The text simply stops — often mid-string or mid-number, with no closing }/].
• If you have access to the API response, the finish reason says length (or max_tokens) rather than stop. That's the definitive signal.

{"users": [{"id": 1, "name": "Ada"}, {"id": 2, "name": "Gr

The string "Gr never closes; two arrays/objects are still open. This is incomplete, not malformed.

Repairing truncated JSON recovers partial data — you've still lost whatever came after the cut. So prevention matters:

• Raise max_tokens (or the model's output cap) to fit the expected response.
• Reduce what you ask for — request fewer fields, or paginate large lists into multiple calls.
• Check the finish reason in code and re-request when it's length:

if (response.choices[0].finish_reason === "length") {
  // response was cut off — retry with a higher limit or smaller request
}

• Use streaming with a running brace counter if you need to detect truncation as it happens.

Sometimes you can't re-run (cost, latency, the data is "good enough"). To salvage what arrived, close the open structures: finish any open string, then append the missing ] and } in the right order. Doing this by hand is fiddly, so use an automated repair — paste the fragment into the AI JSON Cleaner, which balances brackets and closes strings, then run the result through the JSON Validator to confirm it parses.

Treat the recovered object as possibly incomplete: the last element may be partial, so validate against your expected shape before trusting it — see Validating LLM JSON Against a Schema.

Truncation is one of several ways model output breaks. For markdown fences and chatter, see Cleaning JSON from ChatGPT; for the general repair toolkit, Fixing Broken LLM JSON; and to keep payloads small enough to avoid limits in the first place, Counting JSON Tokens for LLMs.

Why does my LLM JSON stop in the middle? The model reached its output-token limit before finishing. The response is truncated, leaving strings and brackets unclosed.

How do I know if JSON was truncated or just malformed? Truncation throws Unexpected end of JSON input and the API's finish reason is length. A malformed-but-complete document throws Unexpected token at a specific position.

Can I recover truncated JSON? You can recover the part that arrived by closing the open brackets and strings, but anything after the cutoff is gone. Raising the token limit and re-running is more reliable when possible.

How do I prevent truncation? Increase max_tokens, ask for less data per call (paginate large lists), and check the finish reason so you can retry when a response is cut off.