MySQL AI User Guide
The ML_GENERATE routine uses the specified
large language model (LLM) to generate text-based content as a
response for the given natural-language query.
This topic contains the following sections:
mysql>SELECT sys.ML_GENERATE('QueryInNaturalLanguage'[,options]);options: JSON_OBJECT(keyvalue[,keyvalue]...)keyvalue: { 'task', {'generation'|'summarization'} |'model_id', 'LargeLanguageModelID' |'context', 'Context' |'language', 'Language' |'temperature',Temperature|'max_tokens',MaxTokens|'top_k',K|'top_p',P|'repeat_penalty',RepeatPenalty|'frequency_penalty',FrequencyPenalty|'stop_sequences', JSON_ARRAY('StopSequence'[, 'StopSequence'] ...) |'speculative_decoding', {true|false} }
Following are ML_GENERATE parameters:
QueryInNaturalLanguage:
specifies the natural-language query that is passed to the
large language model (LLM) handle.
options: specifies optional parameters
as key-value pairs in JSON format. It can include the
following parameters:
task: specifies the task expected
from the LLM. Default value is
generation. Possible values are:
generation: generates
text-based content.
summarization: generates a
summary for existing text-based content.
model_id: specifies the LLM to use
for the task. Default and possible value is
llama3.2-3b-instruct-v1.
To view the lists of available LLMs, see In-Database LLM.
context: specifies the context to
be used for augmenting the query and guide the text
generation of the LLM. Default value is
NULL.
language: specifies the language to
be used for writing queries, ingesting documents, and
generating the output. To set the value of the
language parameter, use the
two-letter ISO 639-1 code for the
language.
Default value is en.
For possible values, to view the list of supported languages, see Section 5.4, “Supported LLM, Embedding Model, and Languages”.
temperature: specifies a
non-negative float that tunes the degree of randomness
in generation. Lower temperatures mean less random
generations.
Default value is 0 for all LLMs.
Possible values are float values between
0 and 5 for the
In-Database LLM.
It is suggested that:
To generate the same output for a particular
prompt every time you run it, set the temperature
to 0.
To generate a random new statement for a particular prompt every time you run it, increase the temperature.
max_tokens: specifies the maximum
number of tokens to predict per generation using an
estimate of three tokens per word. Default value is
256. Possible values are integer
values between1 and
4096.
top_k: specifies the number of top
most likely tokens to consider for text generation at
each step. Default value is 40,
which means that top 40 most likely tokens are
considered for text generation at each step. Possible
values are integer values between 0
and 32000.
top_p: specifies a number,
p, and ensures that only the most
likely tokens with the sum of probabilities
p are considered for generation at
each step. A higher value of p
introduces more randomness into the output. Default
value is 0.95. Possible values are
float values between 0 and
1.
To disable this method, set to
1.0 or 0.
To eliminate tokens with low likelihood, assign
p a lower value. For example,
if set to 0.1, tokens within
top 10% probability are included.
To include tokens with low likelihood, assign
p a higher value. For example,
if set to 0.9, tokens within
top 90% probability are included.
If you are also specifying the
top_k parameter, the LLM considers
only the top tokens whose probabilities add up to
p percent. It ignores the rest of
the k tokens.
repeat_penalty: assigns a penalty
when a token appears repeatedly. High penalties
encourage less repeated tokens and produce more random
outputs. Default value is 1.1.
Possible values are float values between
0 and 2.
frequency_penalty: assigns a
penalty when a token appears frequently. High
penalties encourage less repeated tokens and produce
more random outputs. Default value is
0. Possible values are float values
between 0 and 1.
stop_sequences: specifies a list of
characters such as a word, a phrase, a newline, or a
period that tells the LLM when to end the generated
output. If you have more than one stop sequence, then
the LLM stops when it reaches any of those sequences.
Default value is NULL.
Generating text-based content in English using the
llama3.2-3b-instruct-v1 model:
mysql> SELECT sys.ML_GENERATE("What is AI?", JSON_OBJECT("task", "generation", "model_id", "llama3.2-3b-instruct-v1", "language", "en"));
Summarizing English text using the
llama3.2-3b-instruct-v1 model:
mysql> SELECT sys.ML_GENERATE(@text, JSON_OBJECT("task", "summarization", "model_id", "llama3.2-3b-instruct-v1", "language", "en"));
Where, @text is set as shown below:
SET @text="Artificial Intelligence (AI) is a rapidly growing field that has the potential to revolutionize how we live and work. AI refers to the development of computer systems that can perform tasks that typically require human intelligence, such as visual perception, speech recognition, decision-making, and language translation.\n\nOne of the most significant developments in AI in recent years has been the rise of machine learning, a subset of AI that allows computers to learn from data without being explicitly programmed. Machine learning algorithms can analyze vast amounts of data and identify patterns, making them increasingly accurate at predicting outcomes and making decisions.\n\nAI is already being used in a variety of industries, including healthcare, finance, and transportation. In healthcare, AI is being used to develop personalized treatment plans for patients based on their medical history and genetic makeup. In finance, AI is being used to detect fraud and make investment recommendations. In transportation, AI is being used to develop self-driving cars and improve traffic flow.\n\nDespite the many benefits of AI, there are also concerns about its potential impact on society. Some worry that AI could lead to job displacement, as machines become more capable of performing tasks traditionally done by humans. Others worry that AI could be used for malicious ";