general

kjarni_init

enum KjarniErrorCode kjarni_init(void);

Initialize the Kjarni runtime. Optional - auto-initialized on first use.

kjarni_shutdown

void kjarni_shutdown(void);

Shutdown and cleanup. Call before process exit.

kjarni_version

const char *kjarni_version(void);

Get the library version string.

kjarni_float_array_free

void kjarni_float_array_free(struct KjarniFloatArray arr);

Free a float array allocated by Kjarni.

Safety

• arr must be a value returned by a search/embedding call, passed here exactly once. Its buffers are invalid to read afterwards.

kjarni_float_2d_array_free

void kjarni_float_2d_array_free(struct KjarniFloat2DArray arr);

Free a 2D float array allocated by Kjarni.

Safety

• arr must be a value returned by kjarni_embedder_encode_batch, passed here exactly once. Its buffers are invalid to read afterwards.

kjarni_string_free

void kjarni_string_free(char *s);

Free a string allocated by Kjarni.

Safety

• s must be a valid, writable pointer to a std::ffi::c_char. It is written on both success and failure, and any buffers it receives are owned by the caller.

kjarni_string_array_free

void kjarni_string_array_free(struct KjarniStringArray arr);

Free a string array allocated by Kjarni.

Safety

• arr must be a value returned by a call that returns one, passed here exactly once. Its buffers are invalid to read afterwards.

kjarni_cosine_similarity

float kjarni_cosine_similarity(const float *a, const float *b, uintptr_t len);

Safety

• a must be null, or point to at least len readable f32 values.
• b must be null, or point to at least len readable f32 values.

kjarni_generation_config_default

struct KjarniGenerationConfig kjarni_generation_config_default(void);

Get default generation configuration (all sentinels = use model defaults).

kjarni_class_results_free

void kjarni_class_results_free(struct KjarniClassResults results);

Safety

• results must be a value returned by kjarni_classifier_classify, passed here exactly once. Its buffers are invalid to read afterwards.

kjarni_class_batch_free

void kjarni_class_batch_free(struct KjarniClassBatch batch);

Frees a batch from kjarni_classifier_classify_batch and every result in it.

Safety

batch must be a value returned by kjarni_classifier_classify_batch, passed here exactly once.

kjarni_error_code_to_string

const char *kjarni_error_code_to_string(enum KjarniErrorCode err);

kjarni_error_name

const char *kjarni_error_name(enum KjarniErrorCode err);

Get the name of an error code as a C string.

kjarni_last_error_message

const char *kjarni_last_error_message(void);

Get the last error message. Returns NULL if no error.

kjarni_clear_error

void kjarni_clear_error(void);

Clear the last error.

kjarni_sigmoid

float kjarni_sigmoid(float x);

Puts a cross-encoder logit on a 0..1 scale.

Exported rather than reimplemented per binding. It is one line of arithmetic, but the branch is not decoration: exp(-x) overflows to infinity for a large negative x and inf / inf is NaN, which sorts unpredictably. A binding author writing the obvious one-liner gets a function that passes every value they try and returns NaN in production, so there is one implementation and it lives here.

Monotonic, so ranking is identical either way; only the numbers move. Scores stay logits by default because that is what PyTorch returns.

kjarni_rerank_results_free

void kjarni_rerank_results_free(struct KjarniRerankResults results);

Free rerank results

Safety

• results must be a value returned by a rerank call, passed here exactly once. Its buffers are invalid to read afterwards.

kjarni_transcription_free

void kjarni_transcription_free(struct KjarniTranscription transcription);

Free a transcript and every string in it.

Safety

• transcription must be a value written by a transcribe call, passed here exactly once.