language_model.h File Reference

#include <groonga/progress.h>

Include dependency graph for language_model.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Macros
#define	GRN_LANGUAGE_MODEL_LOADER_N_GPU_LAYERS_DEFAULT   999
	The default N GPU layers to use in language model. In general, we use GPU as much as possible by default.

Typedefs
typedef struct grn_language_model_	grn_language_model
	Language model.
typedef struct grn_language_model_inferencer_	grn_language_model_inferencer
	Language model inferencer.
typedef struct grn_language_model_loader_	grn_language_model_loader
	Language model loader.

Functions
GRN_API grn_language_model_loader *	grn_language_model_loader_open (grn_ctx *ctx)
	Open a new language model loader.
GRN_API grn_rc	grn_language_model_loader_close (grn_ctx *ctx, grn_language_model_loader *loader)
	Close a language model loader.
GRN_API grn_rc	grn_language_model_loader_set_model (grn_ctx *ctx, grn_language_model_loader *loader, const char *model, int64_t model_length)
	Set language model name to load.
GRN_API grn_rc	grn_language_model_loader_set_n_gpu_layers (grn_ctx *ctx, grn_language_model_loader *loader, int32_t n_gpu_layers)
	Set the number of GPU layers to use.
GRN_API grn_language_model *	grn_language_model_loader_load (grn_ctx *ctx, grn_language_model_loader *loader)
	Load a language model.
GRN_API uint32_t	grn_language_model_get_n_embedding_dimensions (grn_ctx *ctx, grn_language_model *model)
	Get the number of dimensions of embedding generated by this model.
GRN_API grn_rc	grn_language_model_close (grn_ctx *ctx, grn_language_model *model)
	Close a language model.
GRN_API grn_language_model_inferencer *	grn_language_model_open_inferencer (grn_ctx *ctx, grn_language_model *model)
	Open a new language model inferencer.
GRN_API grn_rc	grn_language_model_inferencer_close (grn_ctx *ctx, grn_language_model_inferencer *inferencer)
	Close a language model inferencer.
GRN_API grn_rc	grn_language_model_inferencer_set_input_column_value_prefix (grn_ctx *ctx, grn_language_model_inferencer *inferencer, const char *prefix, int64_t prefix_length)
	Prepend prefix to all values of input_column in grn_language_model_inferencer_vectorize_in_batch and grn_language_model_inferencer_vectorize_applier.
GRN_API grn_rc	grn_language_model_inferencer_set_progress_callback (grn_ctx *ctx, grn_language_model_inferencer *inferencer, grn_progress_callback_func callback, void *user_data)
	Set progress callback that is called on each batch vectorization is completed.
GRN_API grn_rc	grn_language_model_inferencer_vectorize (grn_ctx *ctx, grn_language_model_inferencer *inferencer, const char *text, int64_t text_length, grn_obj *output_vector)
	Vectorize a text.
GRN_API grn_rc	grn_language_model_inferencer_vectorize_applier (grn_ctx *ctx, grn_language_model_inferencer *inferencer, grn_obj *input_column, grn_applier_data *data)
	Vectorize texts in batch.
GRN_API grn_rc	grn_language_model_inferencer_vectorize_in_batch (grn_ctx *ctx, grn_language_model_inferencer *inferencer, grn_table_cursor *cursor, grn_obj *input_column, grn_obj *output)
	Vectorize texts and output embeddings set to Float32 GRN_UVECTOR or vector column in batch.

Macro Definition Documentation

GRN_LANGUAGE_MODEL_LOADER_N_GPU_LAYERS_DEFAULT

#define GRN_LANGUAGE_MODEL_LOADER_N_GPU_LAYERS_DEFAULT   999

The default N GPU layers to use in language model. In general, we use GPU as much as possible by default.

This value is same as n_gpu_layers of llama_model_default_params().

Since

15.2.1

Typedef Documentation

grn_language_model

typedef struct grn_language_model_ grn_language_model

Language model.

   You need to use \ref grn_language_model_loader_load to load
   \ref grn_language_model.

grn_language_model_inferencer

typedef struct grn_language_model_inferencer_ grn_language_model_inferencer

Language model inferencer.

   You need to use \ref grn_language_model_open_inferencer to
   open \ref grn_language_model_inferencer.

grn_language_model_loader

typedef struct grn_language_model_loader_ grn_language_model_loader

Language model loader.

Function Documentation

grn_language_model_close()

GRN_API grn_rc grn_language_model_close	(	grn_ctx *	ctx,
		grn_language_model *	model
	)

Close a language model.

Parameters

ctx	The context object.
model	The model to close.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

grn_language_model_get_n_embedding_dimensions()

GRN_API uint32_t grn_language_model_get_n_embedding_dimensions	(	grn_ctx *	ctx,
		grn_language_model *	model
	)

Get the number of dimensions of embedding generated by this model.

Parameters

ctx	The context object.
model	The model.

Returns

The number of dimensions on success, 0 on error.

    See `ctx->rc` for error details.

grn_language_model_inferencer_close()

GRN_API grn_rc grn_language_model_inferencer_close	(	grn_ctx *	ctx,
		grn_language_model_inferencer *	inferencer
	)

Close a language model inferencer.

Parameters

ctx	The context object.
inferencer	The inferencer to close.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

grn_language_model_inferencer_set_input_column_value_prefix()

GRN_API grn_rc grn_language_model_inferencer_set_input_column_value_prefix	(	grn_ctx *	ctx,
		grn_language_model_inferencer *	inferencer,
		const char *	prefix,
		int64_t	prefix_length
	)

Prepend prefix to all values of input_column in grn_language_model_inferencer_vectorize_in_batch and grn_language_model_inferencer_vectorize_applier.

Parameters

ctx	The context object.
inferencer	The inferencer.
prefix	The prefix.
prefix_length	The byte size of prefix. You can use -1 if prefix is a \0-terminated string.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

Since

15.1.9

grn_language_model_inferencer_set_progress_callback()

GRN_API grn_rc grn_language_model_inferencer_set_progress_callback	(	grn_ctx *	ctx,
		grn_language_model_inferencer *	inferencer,
		grn_progress_callback_func	callback,
		void *	user_data
	)

Set progress callback that is called on each batch vectorization is completed.

Parameters

ctx	The context object.
inferencer	The inferencer.
callback	The callback.
user_data	The data that is passed to the callback when callback is called.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

Since

15.2.1

grn_language_model_inferencer_vectorize()

GRN_API grn_rc grn_language_model_inferencer_vectorize	(	grn_ctx *	ctx,
		grn_language_model_inferencer *	inferencer,
		const char *	text,
		int64_t	text_length,
		grn_obj *	output_vector
	)

Vectorize a text.

In other words, compute embeddings of a text.

Parameters

ctx	The context object.
inferencer	The inferencer.
text	The text to vectorize.
text_length	The byte size of text. You can use -1 if text is a \0-terminated string.
output_vector	GRN_DB_FLOAT32 vector as an output. A caller must initialize this by GRN_FLOAT32_INIT and GRN_OBJ_VECTOR.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

grn_language_model_inferencer_vectorize_applier()

GRN_API grn_rc grn_language_model_inferencer_vectorize_applier	(	grn_ctx *	ctx,
		grn_language_model_inferencer *	inferencer,
		grn_obj *	input_column,
		grn_applier_data *	data
	)

Vectorize texts in batch.

In other words, compute embeddings set of texts. This is efficient than calling grn_language_model_inferencer_vectorize() multiple times.

This should be used from an applier.

Parameters

ctx	The context object.
inferencer	The inferencer.
input_column	The text family column or accessor. Caller must be ensure it. This function doesn't validate it.
data	The applier data passed to an applier function.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

grn_language_model_inferencer_vectorize_in_batch()

GRN_API grn_rc grn_language_model_inferencer_vectorize_in_batch	(	grn_ctx *	ctx,
		grn_language_model_inferencer *	inferencer,
		grn_table_cursor *	cursor,
		grn_obj *	input_column,
		grn_obj *	output
	)

Vectorize texts and output embeddings set to Float32 GRN_UVECTOR or vector column in batch.

In other words, compute embeddings set of texts. This is efficient than calling grn_language_model_inferencer_vectorize() multiple times.

You can use grn_memory_map to write and read the result embeddings set without allocating memory for all embeddings set:

uint32_t n_dimensions = 256;
uint32_t n_records = 1000;
size_t embeddings_set_size = n_dimensions * n_records;
grn_memory_map *embeddings_set_map =
  grn_memory_map_open(ctx,
                      "/tmp/embeddings",
                      GRN_MEMORY_MAP_READ | GRN_MEMORY_MAP_WRITE,
                      0,
                      embeddings_set_size);
float *embeddings_set_raw =
  grn_memory_map_get_address(ctx, embeddings_set_map);
grn_obj embeddings_set;
GRN_FLOAT32_INIT(&embeddings_set, GRN_OBJ_VECTOR | GRN_OBJ_DO_SHALLOW_COPY);
GRN_BINARY_SET_REF(&embeddings_set, embeddings_set_raw, embeddings_set_size);
GRN_BULK_REWIND(&embeddings_set);
grn_table_cursor *cursor =
  grn_table_cursor_open(ctx,
                        source_table,
                        NULL, 0,
                        NULL, 0,
                        0, -1, GRN_CURSOR_BY_ID);
if (cursor) {
  grn_language_model_inferencer_vectorize_in_batch(ctx,
                                                   inferencer,
                                                   cursor,
                                                   source_column,
                                                   &embeddings_set);
  grn_table_cursor_close(ctx, cursor);
}
// Use &embedding_set
GRN_OBJ_FIN(ctx, &embedding_set);
grn_memory_map_close(ctx, embeddings_set_map);

Parameters

ctx	The context object.
inferencer	The inferencer.
cursor	The cursor that returns target record IDs.
input_column	The text family column or accessor. Caller must be ensure it. This function doesn't validate it.
output	The generated embeddings set. This must be a Float32 vector or Float32 vector column. Output order is same as IDs returned by the cursor.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

grn_language_model_loader_close()

GRN_API grn_rc grn_language_model_loader_close	(	grn_ctx *	ctx,
		grn_language_model_loader *	loader
	)

Close a language model loader.

Parameters

ctx	The context object.
loader	The loader to close.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

grn_language_model_loader_load()

GRN_API grn_language_model * grn_language_model_loader_load	(	grn_ctx *	ctx,
		grn_language_model_loader *	loader
	)

Load a language model.

   If the target model is already loaded, this reuses the model
   instead of loading a new model.

Parameters

ctx	The context object.
loader	The loader.

Returns

A loaded grn_language_model on success, NULL on error.

    See `ctx->rc` for error details.

grn_language_model_loader_open()

GRN_API grn_language_model_loader * grn_language_model_loader_open

(

grn_ctx *

ctx

)

Open a new language model loader.

Parameters

ctx	The context object.

Returns

A newly created language model loader on success, NULL on error.

grn_language_model_loader_set_model()

GRN_API grn_rc grn_language_model_loader_set_model	(	grn_ctx *	ctx,
		grn_language_model_loader *	loader,
		const char *	model,
		int64_t	model_length
	)

Set language model name to load.

Parameters

ctx	The context object.
loader	The loader.
model	The model name to load.
model_length	The byte size of model. You can use -1 if model is a \0-terminated string.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

Since

15.1.8

grn_language_model_loader_set_n_gpu_layers()

GRN_API grn_rc grn_language_model_loader_set_n_gpu_layers	(	grn_ctx *	ctx,
		grn_language_model_loader *	loader,
		int32_t	n_gpu_layers
	)

Set the number of GPU layers to use.

You can disable GPU by specifying 0 as n_gpu_layers.

Parameters

ctx	The context object.
loader	The loader.
n_gpu_layers	The number of GPU layers to use.

Returns

GRN_SUCCESS on success, the appropriate grn_rc on error.

Since

15.2.1

grn_language_model_open_inferencer()

GRN_API grn_language_model_inferencer * grn_language_model_open_inferencer	(	grn_ctx *	ctx,
		grn_language_model *	model
	)

Open a new language model inferencer.

Parameters

ctx	The context object.
model	The model to inference.

Returns

A newly created grn_language_model_inferencer on success, NULL on error.

See ctx->rc for error details.