85 lines
2.6 KiB
C++
85 lines
2.6 KiB
C++
#include "google/optimize/model_json.h"
|
|||
|
|
|
||
|
|
#include "google/errors.h"
|
||
|
|
#include "google/optimize/caps.h"
|
||
|
|
|
||
|
|
namespace molo::google::optimize {
|
||
|
|
|
||
|
|
using clients::body::Json;
|
||
|
|
|
||
|
|
JsonModelCall::JsonModelCall(std::shared_ptr<LanguageModel> model,
|
||
|
|
std::string model_name)
|
||
|
|
: model_(std::move(model)),
|
||
|
|
model_name_(std::move(model_name)),
|
||
|
|
calls_(std::make_shared<int>(0)) {
|
||
|
|
if (!model_) throw GoogleError::invalid("a language model is required");
|
||
|
|
}
|
||
|
|
|
||
|
|
GenerateRequest JsonModelCall::build(const std::string& system,
|
||
|
|
const std::string& user, bool json) const {
|
||
|
|
if (user.size() > OptimizeCaps::kMaxTranscriptChars + OptimizeCaps::kMaxTextChars)
|
||
|
|
throw GoogleError::invalid("model input over the character cap");
|
||
|
|
GenerateRequest req;
|
||
|
|
req.model = model_name_;
|
||
|
|
if (!system.empty()) {
|
||
|
|
Content sys;
|
||
|
|
sys.parts.push_back(Part::of_text(system));
|
||
|
|
req.system = sys;
|
||
|
|
}
|
||
|
|
req.contents.push_back(Content::user(user));
|
||
|
|
req.config.temperature = 0.0;
|
||
|
|
req.config.max_output_tokens = 1024;
|
||
|
|
req.config.thinking_budget = 0;
|
||
|
|
if (json) req.config.response_mime_type = "application/json";
|
||
|
|
return req;
|
||
|
|
}
|
||
|
|
|
||
|
|
Json JsonModelCall::extract(const std::string& text) {
|
||
|
|
std::size_t start = text.find('{');
|
||
|
|
while (start != std::string::npos) {
|
||
|
|
int depth = 0;
|
||
|
|
bool in_str = false;
|
||
|
|
bool esc = false;
|
||
|
|
for (std::size_t i = start; i < text.size(); ++i) {
|
||
|
|
char c = text[i];
|
||
|
|
if (in_str) {
|
||
|
|
if (esc) esc = false;
|
||
|
|
else if (c == '\\') esc = true;
|
||
|
|
else if (c == '"') in_str = false;
|
||
|
|
continue;
|
||
|
|
}
|
||
|
|
if (c == '"') in_str = true;
|
||
|
|
else if (c == '{') ++depth;
|
||
|
|
else if (c == '}' && --depth == 0) {
|
||
|
|
try {
|
||
|
|
Json doc = Json::parse(text.substr(start, i - start + 1));
|
||
|
|
if (doc.is_object()) return doc;
|
||
|
|
} catch (const clients::body::JsonError&) {
|
||
|
|
}
|
||
|
|
break;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
start = text.find('{', start + 1);
|
||
|
|
}
|
||
|
|
return Json();
|
||
|
|
}
|
||
|
|
|
||
|
|
Json JsonModelCall::ask(const std::string& system, const std::string& user) const {
|
||
|
|
GenerateRequest req = build(system, user, true);
|
||
|
|
for (int attempt = 0; attempt <= OptimizeCaps::kMaxModelRetries; ++attempt) {
|
||
|
|
++*calls_;
|
||
|
|
Json doc = extract(model_->generate(req).text());
|
||
|
|
if (doc.is_object()) return doc;
|
||
|
|
}
|
||
|
|
throw GoogleError(502, "BAD_RESPONSE", "model reply held no JSON object",
|
||
|
|
"optimize");
|
||
|
|
}
|
||
|
|
|
||
|
|
std::string JsonModelCall::ask_text(const std::string& system,
|
||
|
|
const std::string& user) const {
|
||
|
|
++*calls_;
|
||
|
|
return model_->generate(build(system, user, false)).text();
|
||
|
|
}
|
||
|
|
|
||
|
|
} // namespace molo::google::optimize
|