Gera Tools

JSON Pointer Extractor

Extract values from LLM JSON output using RFC 6901 JSON Pointers.

Free JSON Pointer evaluator. Paste a JSON object and an RFC 6901 pointer like /choices/0/message/content to pull deeply nested fields out of structured LLM responses, all in your browser. It runs free in your browser on Gera Tools, with nothing uploaded.

Last updated Source: Gera Tools

What is a JSON Pointer?

A JSON Pointer (RFC 6901) is a compact string syntax for identifying a specific value inside a JSON document. Segments are separated by slashes, so /a/b/0 means "key a, then key b, then array element 0".

Extract any field from structured LLM output

Models that return JSON — tool calls, function calls, structured outputs — bury the value you actually want deep inside wrapper objects. This tool lets you pull it out with a single RFC 6901 JSON Pointer, the standard, unambiguous way to address a node in a JSON document. Paste the response, type a pointer like /choices/0/message/content, and get the value back instantly.

How JSON Pointer evaluation works

A JSON Pointer is a string of reference tokens, each prefixed by a /. Starting from the root document, the evaluator walks one token at a time: for an object it looks up the token as a key, for an array it treats the token as a numeric index. The empty pointer "" refers to the whole document. Two escape sequences keep the syntax unambiguous — ~1 means a literal / inside a key, and ~0 means a literal ~. If any token fails to resolve (missing key, out-of-range index), the whole pointer has no match.

Tips and examples

  • OpenAI chat completions: /choices/0/message/content grabs the assistant text.
  • Tool/function arguments often arrive as a JSON string — extract it, then paste that string back in as a fresh document to drill deeper.
  • To pull every element of an array you address them one index at a time; a pointer resolves to exactly one node by design.
  • If a key literally contains a slash (rare, but valid JSON), escape it as ~1.

JSON Pointer vs JSONPath — which to use?

Both are standard ways to address values inside a JSON document, but they have different design goals:

JSON Pointer (RFC 6901) resolves to exactly one node. The path is compact and unambiguous — there are no wildcards, no filters, no fanout. Every JSON Pointer either resolves to a single value or fails. This makes it ideal for JSON Patch (RFC 6902), JSON Merge Patch, and OpenAPI discriminator mappings, where a precise, deterministic reference is required.

JSONPath is expressive and can return multiple matches through wildcards ([*]) and recursive descent (..). It is the right choice when you want to fan out over an array — for example, extracting every id from a list of objects.

For navigating to one specific field in a structured LLM response, JSON Pointer is cleaner and less ambiguous. For collecting all values matching a pattern, use JSONPath.

Common RFC 6901 patterns for LLM output

OpenAI Chat Completions

/choices/0/message/content

The finish reason:

/choices/0/finish_reason

Tool call function name:

/choices/0/message/tool_calls/0/function/name

Anthropic Messages API

/content/0/text

Input token usage:

/usage/input_tokens

Escaping in practice

The two escape sequences exist to handle the characters that are special in the pointer syntax:

Literal character in keyWrite in pointer
~ (tilde)~0
/ (forward slash)~1

For example, to reach the value at key a/b inside the root object, the pointer is /a~1b. Keys containing tildes or slashes are unusual but valid JSON, and the escaping ensures the pointer remains unambiguous.

The empty string "" is a valid JSON Pointer that references the entire document — useful for confirming that your JSON parsed correctly without navigating anywhere.