Implementation:Ggml org Llama cpp Jinja String Header

Knowledge Sources	Ggml_org_Llama_cpp
Domains	Template_Engine
Last Updated	2026-02-15 00:00 GMT

Overview

Declares the `jinja::string_part` and `jinja::string` types that track input provenance through string operations.

Description

The `string_part` struct pairs a `std::string` value with an `is_input` boolean flag and provides case-checking helpers. The `jinja::string` struct wraps a `vector<string_part>` and provides constructors from `std::string`, `int`, and `double`. It declares methods for marking input, converting to plain string, computing length, hashing, checking case, appending, and in-place transformations (uppercase, lowercase, capitalize, titlecase, strip). Transformation rules preserve provenance: one-to-one operations keep the `is_input` flag, one-to-many propagate it, and many-to-one mark as input only if all source parts are input.

Usage

Include this header when working with the Jinja engine's string system. It defines the foundational string type for the input marking security mechanism, enabling safe handling of user input in chat templates by tracking data provenance at the string-part level.

Code Reference

Source Location

• Repository: Ggml_org_Llama_cpp
• File: common/jinja/string.h
• Lines: 1-61

Signature

namespace jinja {

struct string_part {
    bool is_input = false;
    std::string val;
    bool is_uppercase() const;
    bool is_lowercase() const;
};

struct string {
    std::vector<string_part> parts;
    string() = default;
    string(const std::string & v, bool user_input = false);
    string(int v);
    string(double v);

    void mark_input();
    std::string str() const;
    size_t length() const;
    void hash_update(hasher & hash) const noexcept;
    bool all_parts_are_input() const;
    bool is_uppercase() const;
    bool is_lowercase() const;
    void mark_input_based_on(const string & other);
    string append(const string & other);
    string uppercase();
    string lowercase();
    string capitalize();
    string titlecase();
    string strip(bool left, bool right, std::optional<const std::string_view> chars = std::nullopt);
};

} // namespace jinja

Import

#include "jinja/string.h"

I/O Contract

Inputs

Name	Type	Required	Description
v	const std::string &	Yes	String value to construct a jinja::string from
user_input	bool	No	Whether this string originates from user input (default false)
other	const string &	Yes	Another jinja::string for append or mark_input_based_on operations
left	bool	Yes	Whether to strip characters from the left side
right	bool	Yes	Whether to strip characters from the right side
chars	std::optional<const std::string_view>	No	Specific characters to strip; defaults to whitespace

Outputs

Name	Type	Description
str return	std::string	Plain concatenated string from all parts
length return	size_t	Total character count across all parts
all_parts_are_input return	bool	True if every part has is_input set
transformation returns	string	New jinja::string with provenance flags preserved through the transformation

Usage Examples

#include "jinja/string.h"

// Create a string tracking user input provenance
jinja::string_part part{true, "user message"};  // is_input = true

jinja::string s("template text");  // is_input = false by default
jinja::string user_s("user text", true);  // is_input = true

// Check provenance
bool from_user = user_s.all_parts_are_input();  // true

// Transform while preserving provenance
auto upper = user_s.uppercase();  // still marked as input

Related Pages

• Principle:Ggml_org_Llama_cpp_Input_Provenance_Tracking