fix: stabilize Gemma 4 brewery generation

remove misleading turn-token output guidance from the brewery prompt
extract the last balanced JSON object before validation
keep README model setup and run instructions aligned
preserve Gemma 4 sampling defaults and local model usage

pipeline/README.md

@@ -1,8 +1,8 @@
 # Biergarten Pipeline
 
-Biergarten Pipeline is a C++23 command-line tool that reads a local city list, resolves contextual enrichment for each sampled city through an injected service, and generates brewery names and descriptions. The current code samples up to four locations per run, then uses either a local GGUF model or the mock generator to produce the output.
+Biergarten Pipeline is a C++23 command-line tool that reads a local city list, resolves contextual enrichment for each sampled city through an injected service, and generates brewery names and descriptions. The current code samples up to four locations per run, then uses either Gemma 4 or the mock generator to produce the output.
 
-## Hardware & GPU Config
+## Tested Hardware & OS
 
 ### x86/64 Linux, NVIDIA RTX 2000
 
@@ -10,7 +10,7 @@ Biergarten Pipeline is a C++23 command-line tool that reads a local city list, r
 - **CPU**: Intel Core Ultra 7 155H
 - **GPU**: NVIDIA RTX 2000 Ada Generation
 - **Memory**: 32GB
-- **Model**: Qwen3-8B-Q6-K
+- **Model**: Gemma 4 E4B: efficient local reasoning; released Apr 2, 2026.
 - **Inference**: llama.cpp with CUDA 12.x support
 
 ### ARM MacOS, M1 Pro
@@ -19,7 +19,7 @@ Biergarten Pipeline is a C++23 command-line tool that reads a local city list, r
 - **CPU**: Apple M1 Pro (8-core)
 - **GPU**: Apple M1 Pro (14-core) [Integrated]
 - **Memory**: 16GB
-- **Model**: gemma-4-E4B-it-Q6_K.gguf
+- **Model**: Gemma 4 E4B: efficient local reasoning; released Apr 2, 2026.
 - **Inference**: llama.cpp with Metal (MPS) support
 
 ## Pipeline
@@ -54,7 +54,7 @@ If an enrichment lookup throws, the pipeline skips that city and keeps going. If
 | libcurl              | Required for Wikipedia requests.                                           |
 | Optional GPU tooling | CUDA on NVIDIA, HIP/ROCm on supported AMD systems, Metal on Apple Silicon. |
 
-Boost, Boost.DI, spdlog, and llama.cpp are fetched by CMake. On Apple Silicon, Metal is enabled automatically. On Linux, the build looks for CUDA or HIP/ROCm when the matching toolkit is present. Windows is not supported.
+Boost, Boost.DI, spdlog, and llama.cpp are fetched by CMake. On Apple Silicon, Metal is enabled automatically. On Linux, the build looks for CUDA or HIP/ROCm when the matching toolkit is present. There are no plans to support Windows.
 
 ```bash
 cmake -S . -B build
@@ -63,19 +63,30 @@ cmake --build build
 
 If the dependency build fails on macOS, check the repo build notes.
 
+## Model
+
+Create a `models/` directory and download the GGUF file there before running the app.
+
+```bash
+mkdir -p models
+curl -L \
+	-o models/google_gemma-4-E4B-it-Q6_K.gguf \
+	https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF/resolve/main/google_gemma-4-E4B-it-Q6_K.gguf?download=true
+```
+
 ## Run
 
 Run the executable from the build directory so the copied `locations.json` is available.
 
 ```bash
 ./biergarten-pipeline --mocked
-./biergarten-pipeline --model /path/to/model.gguf --temperature 1.0 --top-p 0.95 --top-k 64 --n-ctx 8192 --seed -1
+./biergarten-pipeline --model models/google_gemma-4-E4B-it-Q6_K.gguf --temperature 1.0 --top-p 0.95 --top-k 64 --n-ctx 8192 --seed -1
 ```
 
 | Flag            | Purpose                                                                      |
-| --------------- | -------------------------------------------- |
+| --------------- | ---------------------------------------------------------------------------- |
 | `--mocked`      | Uses the mock generator instead of a model.                                  |
-| `--model, -m`   | Path to a GGUF model file.                   |
+| `--model, -m`   | Path to a GGUF model file, such as `models/google_gemma-4-E4B-it-Q6_K.gguf`. |
 | `--temperature` | Sampling temperature. Default: `1.0`.                                        |
 | `--top-p`       | Nucleus sampling parameter. Default: `0.95`.                                 |
 | `--top-k`       | Top-k sampling parameter. Default: `64`.                                     |

pipeline/includes/data_generation/llama_generator_helpers.h

@@ -77,4 +77,12 @@ std::string ValidateBreweryJsonPublic(const std::string& raw,
                                       std::string& name_out,
                                       std::string& description_out);
 
+/**
+ * @brief Extracts the last balanced JSON object from text.
+ *
+ * @param text Input text.
+ * @return Extracted JSON object or an empty string if none exists.
+ */
+std::string ExtractLastJsonObjectPublic(const std::string& text);
+
 #endif  // BIERGARTEN_PIPELINE_DATA_GENERATION_LLAMA_GENERATOR_HELPERS_H_

pipeline/prompts/system.md

@@ -1,10 +1,7 @@
 <|think|>
-CRITICAL INSTRUCTION: You must use the <|think|> token to reason through the brewery's details before providing the final JSON output. Inside the think block, verify that you are not using blacklisted terms and that your technical and architectural details are unique.
+Think through the brewery details internally before answering.
-
+Return only one raw JSON object as the final answer, with exactly two keys: "name" and "description".
-Structure your response as follows:
+No markdown, code fences, preamble, or extra keys.
-[Your reasoning and constraint checklist go here]
-<|turn|>
-{"name": "...", "description": "..."}
 
 # FULL SYSTEM PROMPT
 
@@ -28,7 +25,7 @@ $$Information about local beer culture, history, or geography$$
 
 ## CRITICAL OUTPUT FORMAT (READ CAREFULLY):
 
-You have to return a reasoning block first, then ONLY raw, perfectly valid JSON after the `<|turn|>` separator. Any mistake with the JSON means the data pipeline breaks.
+You have to return a reasoning block first, then ONLY raw, perfectly valid JSON as the final answer. Any mistake with the JSON means the data pipeline breaks.
 
 ABSOLUTELY NO MARKDOWN FORMATTING. Do NOT wrap your response in json or ``` blocks.
 

pipeline/src/data_generation/llama/generate_brewery.cpp

@@ -26,8 +26,9 @@ auto ExtractFinalJsonPayload(std::string raw_response) -> std::string {
       return text.substr(first, last - first + 1);
    };
 
-   static const std::array<std::string_view, 4> separator_tokens = {
+   static const std::array<std::string_view, 6> separator_tokens = {
-       "<|turn|>", "<turn|>", "<channel|>", "<|channel|>"};
+       "<|think|>", "<think|>",   "<|turn|>",
+       "<turn|>",   "<channel|>", "<|channel|>"};
 
    std::size_t separator_pos = std::string::npos;
    std::size_t separator_length = 0;
@@ -46,20 +47,15 @@ auto ExtractFinalJsonPayload(std::string raw_response) -> std::string {
    }
 
    const std::string_view trimmed = trim(raw_response);
-   const std::size_t first_brace = trimmed.find('{');
+   std::string json_candidate =
-   if (first_brace == std::string_view::npos) {
+       ExtractLastJsonObjectPublic(std::string(trimmed));
+   if (!json_candidate.empty()) {
+      return ExtractLastJsonObjectPublic(std::string(trimmed));
+   }
+
    return std::string(trimmed);
 }
 
-   const std::size_t last_brace = trimmed.find_last_of('}');
-   if (last_brace == std::string_view::npos || last_brace < first_brace) {
-      return std::string(trimmed.substr(first_brace));
-   }
-
-   return std::string(
-       trimmed.substr(first_brace, last_brace - first_brace + 1));
-}
-
 }  // namespace
 
 auto LlamaGenerator::GenerateBrewery(const BreweryLocation& location,
@@ -147,9 +143,11 @@ auto LlamaGenerator::GenerateBrewery(const BreweryLocation& location,
       // limits.
       prompt =
           "Your previous response was invalid. Error: " + validation_error +
-          "\nReturn ONLY valid JSON with this exact schema: "
+          "\nReturn ONLY valid JSON with exactly these keys: "
-          "{\"name\": \"string\", \"description\": \"string\"}."
+          "{\"name\": \"<brewery name>\", "
-          "\nDo not include markdown, comments, or extra keys.";
+          "\"description\": \"<single-paragraph description>\"}."
+          "\nDo not include markdown, comments, extra keys, or literal "
+          "placeholder values.";
       prompt += "\n\n";
       prompt += retry_location;
    }

pipeline/src/data_generation/llama/helpers.cpp

@@ -263,12 +263,14 @@ static void AppendTokenPiece(const llama_vocab* vocab, llama_token token,
    output.append(buffer.data(), static_cast<std::size_t>(bytes));
 }
 
-static bool ExtractFirstJsonObject(const std::string& text,
+static bool ExtractLastJsonObject(const std::string& text,
                                   std::string& json_out) {
    std::size_t start = std::string::npos;
    int depth = 0;
    bool in_string = false;
    bool escaped = false;
+   bool found = false;
+   std::string candidate;
 
    for (std::size_t i = 0; i < text.size(); ++i) {
       const char ch = text[i];
@@ -303,15 +305,29 @@ static bool ExtractFirstJsonObject(const std::string& text,
          }
          --depth;
          if (depth == 0 && start != std::string::npos) {
-            json_out = text.substr(start, i - start + 1);
+            candidate = text.substr(start, i - start + 1);
-            return true;
+            found = true;
          }
       }
    }
 
+   if (!found) {
       return false;
    }
 
+   json_out = std::move(candidate);
+   return true;
+}
+
+std::string ExtractLastJsonObjectPublic(const std::string& text) {
+   std::string extracted;
+   if (ExtractLastJsonObject(text, extracted)) {
+      return extracted;
+   }
+
+   return {};
+}
+
 static std::string ValidateBreweryJson(const std::string& raw,
                                        std::string& name_out,
                                        std::string& description_out) {
@@ -371,7 +387,7 @@ static std::string ValidateBreweryJson(const std::string& raw,
    std::string validation_error;
    if (ec) {
       std::string extracted;
-      if (!ExtractFirstJsonObject(raw, extracted)) {
+      if (!ExtractLastJsonObject(raw, extracted)) {
          return "JSON parse error: " + ec.message();
       }