Implementation:Ggml org Llama cpp Json Partial

Knowledge Sources	Ggml_org_Llama_cpp
Domains	JSON, Parsing
Last Updated	2026-02-15 00:00 GMT

Overview

Implements partial/incomplete JSON parsing with "healing" capability, allowing truncated JSON strings from streaming LLM output to be completed and parsed.

Description

This module uses nlohmann::json's SAX interface via a custom json_error_locator to find where parsing fails in an incomplete JSON string. It maintains a stack of JSON elements (objects, keys, arrays) to track nesting depth. When a parse error is found, it inserts a healing marker and closes all open brackets/braces to produce valid JSON, preserving the marker in the output so the JSON can be re-split at that point later.

Usage

Use this module when parsing streaming or partial JSON outputs from LLM models, particularly for implementing partial tool calls in the OpenAI-compatible API format where the model has not finished generating the full JSON response.

Code Reference

Source Location

• Repository: Ggml_org_Llama_cpp
• File: common/json-partial.cpp
• Lines: 1-324

Signature

bool common_json_parse(
    const std::string & input,
    const std::string & healing_marker,
    common_json & out);

bool common_json_parse(
    std::string::const_iterator & it,
    const std::string::const_iterator & end,
    const std::string & healing_marker,
    common_json & out);

Import

#include "json-partial.h"
#include "log.h"
#include <nlohmann/json.hpp>
#include <string>
#include <regex>

I/O Contract

Inputs

Name	Type	Required	Description
input	std::string	Yes	The potentially incomplete JSON string to parse
it	std::string::const_iterator &	Yes	Iterator-based variant: current position in JSON input
end	std::string::const_iterator	Yes	Iterator-based variant: end of JSON input
healing_marker	std::string	Yes	Marker string inserted at the point where healing occurs, for later re-splitting

Outputs

Name	Type	Description
return	bool	True if parsing (with healing) succeeded, false otherwise
out	common_json &	Output struct containing the parsed JSON and healing metadata

Usage Examples

#include "json-partial.h"

// Partial JSON from streaming LLM output
std::string partial = R"({"name": "get_weather", "arguments": {"location": "San Fr)";
std::string marker = "__HEAL__";
common_json result;

if (common_json_parse(partial, marker, result)) {
    // result contains healed JSON: {"name": "get_weather", "arguments": {"location": "San Fr__HEAL__"}}
    // The marker indicates where the original input was truncated
}

Related Pages

• Principle:Ggml_org_Llama_cpp_StreamingJSONParsing