context_fit.h

Go to the documentation of this file.

// SPDX-License-Identifier: Apache-2.0
#ifndef ENTROPIC_INFERENCE_CONTEXT_FIT_H
#define ENTROPIC_INFERENCE_CONTEXT_FIT_H
 
#include <cstddef>
#include <string>
 
namespace entropic {
 
struct ContextFit {
    int prompt_tokens = 0;       
    int context_length = 0;      
    int system_tokens = 0;       
    int tool_tokens = 0;         
    std::size_t tool_bytes = 0;  
    int tool_count = 0;          
};
 
inline bool context_fit_overflows(const ContextFit& f) {
    return f.context_length > 0 && f.prompt_tokens >= f.context_length;
}
 
inline int context_fit_history_tokens(const ContextFit& f) {
    const int rest = f.prompt_tokens - f.system_tokens - f.tool_tokens;
    return rest > 0 ? rest : 0;
}
 
inline std::string context_fit_largest_contributor(const ContextFit& f) {
    int most = context_fit_history_tokens(f);
    std::string who = "the message history";
    if (f.tool_tokens > most) {
        most = f.tool_tokens;
        who = "the staged tool block";
    }
    if (f.system_tokens > most) {
        who = "the system prompt";
    }
    return who;
}
 
inline std::string context_fit_breakdown(const ContextFit& f) {
    return "staged tools " + std::to_string(f.tool_tokens) + " tok ("
         + std::to_string(f.tool_count) + " tools, "
         + std::to_string(f.tool_bytes) + " bytes); system prompt "
         + std::to_string(f.system_tokens) + " tok; message history "
         + std::to_string(context_fit_history_tokens(f)) + " tok";
}
 
inline std::string context_overflow_message(const ContextFit& f) {
    return "Prompt does not fit this tier's context: "
         + std::to_string(f.prompt_tokens) + " prompt tokens against "
           "context_length=" + std::to_string(f.context_length)
         + " (over by "
         + std::to_string(f.prompt_tokens - f.context_length)
         + " tokens). Largest contributor: "
         + context_fit_largest_contributor(f) + ". Breakdown: "
         + context_fit_breakdown(f)
         + ". Refused before decode — compaction cannot shrink the staged "
           "tool block. Raise this tier's context_length, or narrow its "
           "allowed_tools.";
}
 
}  // namespace entropic
 
#endif  // ENTROPIC_INFERENCE_CONTEXT_FIT_H