fix: llama backend lifetime, Wikipedia enrichment depth, and misc cleanup

pipeline/includes/data_generation/llama_generator.h

@@ -7,6 +7,7 @@
  */
 
 #include <cstdint>
+#include <random>
 #include <string>
 
 #include "data_generation/data_generator.h"
@@ -114,7 +115,7 @@ class LlamaGenerator final : public DataGenerator {
    llama_context* context_ = nullptr;
    float sampling_temperature_ = 0.8f;
    float sampling_top_p_ = 0.92f;
-   uint32_t sampling_seed_ = 0xFFFFFFFFu;
+   std::mt19937 rng_;
    uint32_t n_ctx_ = 8192;
    std::string brewery_system_prompt_;
 };

pipeline/includes/data_generation/llama_generator_helpers.h

@@ -21,7 +21,7 @@ typedef int llama_token;
  * @return Processed region context.
  */
 std::string PrepareRegionContextPublic(std::string_view region_context,
-                                       std::size_t max_chars = 700);
+                                       std::size_t max_chars = 2000);
 
 /**
  * @brief Parses a response expected to contain two logical lines.

pipeline/includes/llama_backend_state.h

@@ -0,0 +1,32 @@
+#ifndef BIERGARTEN_PIPELINE_LLAMA_BACKEND_STATE_H_
+#define BIERGARTEN_PIPELINE_LLAMA_BACKEND_STATE_H_
+
+/**
+ * @file llama_backend_state.h
+ * @brief RAII guard for llama.cpp backend process lifetime.
+ */
+
+#include <llama.h>
+
+/**
+ * @brief RAII wrapper for llama_backend_init and llama_backend_free.
+ *
+ * Create one instance in application startup before using llama.cpp and keep
+ * it alive for application lifetime.
+ */
+class LlamaBackendState {
+  public:
+   /// @brief Initializes global llama backend state.
+   LlamaBackendState() { llama_backend_init(); }
+
+   /// @brief Cleans up global llama backend state.
+   ~LlamaBackendState() { llama_backend_free(); }
+
+   /// @brief Non-copyable type.
+   LlamaBackendState(const LlamaBackendState&) = delete;
+
+   /// @brief Non-copyable type.
+   LlamaBackendState& operator=(const LlamaBackendState&) = delete;
+};
+
+#endif  // BIERGARTEN_PIPELINE_LLAMA_BACKEND_STATE_H_

pipeline/includes/services/wikipedia_service.h

@@ -24,9 +24,10 @@ class WikipediaService final : public IEnrichmentService {
    [[nodiscard]] std::string GetLocationContext(const Location& loc) override;
 
   private:
-   std::string FetchExtract(std::string_view query) const;
+   std::string FetchExtract(std::string_view query);
    std::shared_ptr<WebClient> client_;
    std::unordered_map<std::string, std::string> cache_;
+   std::unordered_map<std::string, std::string> extract_cache_;
 };
 
 #endif  // BIERGARTEN_PIPELINE_WIKIPEDIA_SERVICE_H_