AaronPo/the-biergarten-app
1
0
Fork 0
mirror of https://github.com/aaronpo97/the-biergarten-app.git synced 2026-06-01 01:54:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enhance ValidateBreweryJson to include reasoning output and update GenerateBrewery to use user_prompt

Browse Source 
Add gemma parser
This commit is contained in:
Aaron Po
2026-04-16 20:06:36 -04:00
parent 44a74ed2ad
commit fcc7a5dc8b
12 changed files with 144 additions and 122 deletions
Download Patch File Download Diff File Expand all files Collapse all files

119
pipeline/src/data_generation/llama/helpers.cc
View File

@@ -4,8 +4,6 @@
* parsing, token decoding, and JSON validation helpers for Llama modules.
*/
#include <spdlog/spdlog.h>
#include <algorithm>
#include <array>
#include <boost/json.hpp>
@@ -81,89 +79,6 @@ std::string PrepareRegionContext(std::string_view region_context,
return normalized;
}
std::string ToChatPrompt(const llama_model* model,
const std::string& system_prompt,
const std::string& user_prompt) {
std::string combined_prompt =
std::format("{}\n\n{}", system_prompt, user_prompt);
const char* template_str = llama_model_chat_template(model, nullptr);
// If metadata is missing (nullptr), attempt to use the built-in "gemma" alias
// to leverage the library's interleaved template for Gemma 4 support.
if (template_str == nullptr) {
template_str = "gemma";
spdlog::info(
"LlamaGenerator: model chat template metadata missing; attempting "
"built-in 'gemma' alias");
}
const std::array<llama_chat_message, 2> messages = {{
{.role = "system", .content = system_prompt.c_str()},
{.role = "user", .content = user_prompt.c_str()},
}};
constexpr std::size_t min_template_buffer_size = 1024;
std::vector<char> buffer(
std::max<std::size_t>(min_template_buffer_size,
(system_prompt.size() + user_prompt.size()) * 4));
auto apply_template_with_resize = [&](const char* tmpl,
const llama_chat_message* chat_messages,
int32_t message_count) -> int32_t {
int32_t result = llama_chat_apply_template(
tmpl, chat_messages, message_count, true, buffer.data(),
static_cast<int32_t>(buffer.size()));
if (result < 0) {
return result;
}
const auto buffer_size = static_cast<int32_t>(buffer.size());
if (result >= buffer_size) {
buffer.resize(static_cast<std::size_t>(result) + 1);
result = llama_chat_apply_template(
tmpl, chat_messages, message_count, true, buffer.data(),
static_cast<int32_t>(buffer.size()));
}
return result;
};
int32_t template_result =
apply_template_with_resize(template_str, messages.data(), 2);
if (template_result >= 0) {
return {buffer.data(), static_cast<size_t>(template_result)};
}
spdlog::warn(
"LlamaGenerator: chat template rejected system/user messages (result "
"{}); trying single user fallback",
template_result);
// FALLBACK: If the template fails (e.g., model rejecting the "system" role),
// combine the system and user prompts into a single "user" message.
const std::array<llama_chat_message, 1> fallback_msg = {{
{.role = "user", .content = combined_prompt.c_str()},
}};
template_result =
apply_template_with_resize(template_str, fallback_msg.data(), 1);
// Ultimate fallback: if GGUF template parsing still fails, use raw text.
if (template_result < 0) {
spdlog::warn(
"LlamaGenerator: chat template fallback failed (result {}); using "
"raw prompt text",
template_result);
return combined_prompt;
}
return {buffer.data(), static_cast<size_t>(template_result)};
}
void AppendTokenPiece(const llama_vocab* vocab, llama_token token,
std::string& output) {
constexpr size_t initial_buffer_size = 256;
@@ -193,6 +108,7 @@ void AppendTokenPiece(const llama_vocab* vocab, llama_token token,
if (!buffer_too_small(bytes)) {
output.append(dynamic_buffer.data(), static_cast<size_t>(bytes));
return;
}
throw std::runtime_error(
@@ -201,7 +117,8 @@ void AppendTokenPiece(const llama_vocab* vocab, llama_token token,
std::optional<std::string> ValidateBreweryJson(const std::string& raw,
std::string& name_out,
std::string& description_out) {
std::string& description_out,
std::string& reasoning_out) {
auto validate_object = [&](const boost::json::value& json_value,
std::string& error_out) -> bool {
if (!json_value.is_object()) {
@@ -209,7 +126,14 @@ std::optional<std::string> ValidateBreweryJson(const std::string& raw,
return false;
}
const auto& obj = json_value.get_object();
if (!obj.contains("reasoning") || !obj.at("reasoning").is_string()) {
error_out = "JSON field 'reasoning' is missing or not a string";
return false;
}
if (!obj.contains("name") || !obj.at("name").is_string()) {
error_out = "JSON field 'name' is missing or not a string";
return false;
@@ -219,6 +143,12 @@ std::optional<std::string> ValidateBreweryJson(const std::string& raw,
error_out = "JSON field 'description' is missing or not a string";
return false;
}
const auto& reasoning_value = obj.at("reasoning").as_string();
reasoning_out = Trim(std::string_view(reasoning_value.data(), reasoning_value.size()));
if (reasoning_out.empty()) {
error_out = "JSON field 'reasoning' must not be empty";
return false;
}
const auto& name_value = obj.at("name").as_string();
const auto& description_value = obj.at("description").as_string();
@@ -239,15 +169,16 @@ std::optional<std::string> ValidateBreweryJson(const std::string& raw,
std::string name_lower = name_out;
std::string description_lower = description_out;
std::ranges::transform(name_lower, name_lower.begin(),
[](unsigned char character) {
return static_cast<char>(std::tolower(character));
});
std::ranges::transform(description_lower, description_lower.begin(),
[](unsigned char character) {
return static_cast<char>(std::tolower(character));
});
auto string_to_lower = [](std::string& str_out) {
std::ranges::transform(str_out, str_out.begin(),
[](unsigned char character) {
return static_cast<char>(std::tolower(character));
});
};
string_to_lower(name_lower);
string_to_lower(description_lower);
if (name_lower == "string" || description_lower == "string") {
error_out = "JSON appears to be a schema placeholder, not content";