Type: feature-request
Description:
When using complex scenarios, we need to teach the LLM how to generate the
response. This technique is a few-shot prompting, where a few messages are part
of the prompt before the human question.
Ollama implemented this using the message array, where the end user define it:
https://github.com/ollama/ollama/blob/dc6fe820512d1046f3a342e384baa64b8ce1758c/docs/api.md?plain=1#L451-L457
I think that in this case, it'll be cool to use something similar which can be appended here:
|
"LLMRequest": { |
|
"properties": { |
|
"query": { |
|
"type": "string", |
|
"title": "Query" |
|
}, |
|
"conversation_id": { |
|
"anyOf": [ |
|
{ |
|
"type": "string" |
|
}, |
|
{ |
|
"type": "null" |
|
} |
|
], |
|
"title": "Conversation Id" |
|
}, |
|
"provider": { |
|
"anyOf": [ |
|
{ |
|
"type": "string" |
|
}, |
|
{ |
|
"type": "null" |
|
} |
|
], |
|
"title": "Provider" |
|
}, |
|
"model": { |
|
"anyOf": [ |
|
{ |
|
"type": "string" |
|
}, |
|
{ |
|
"type": "null" |
|
} |
|
], |
|
"title": "Model" |
|
}, |
|
"attachments": { |
|
"anyOf": [ |
|
{ |
|
"items": { |
|
"$ref": "#/components/schemas/Attachment" |
|
}, |
|
"type": "array" |
|
}, |
|
{ |
|
"type": "null" |
|
} |
|
], |
|
"title": "Attachments" |
|
} |
|
}, |
|
"additionalProperties": false, |
|
"type": "object", |
|
"required": [ |
|
"query" |
|
], |
|
"title": "LLMRequest", |
|
"description": "Model representing a request for the LLM (Language Model) send into OLS service.\n\nAttributes:\n query: The query string.\n conversation_id: The optional conversation ID (UUID).\n provider: The optional provider.\n model: The optional model.\n attachments: The optional attachments.\n\nExample:\n ```python\n llm_request = LLMRequest(query=\"Tell me about Kubernetes\")\n ```", |
|
"examples": [ |
|
{ |
|
"attachments": [ |
|
{ |
|
"attachment_type": "log", |
|
"content": "this is attachment", |
|
"content_type": "text/plain" |
|
}, |
|
{ |
|
"attachment_type": "configuration", |
|
"content": "kind: Pod\n metadata:\n name: private-reg", |
|
"content_type": "application/yaml" |
|
}, |
|
{ |
|
"attachment_type": "configuration", |
|
"content": "foo: bar", |
|
"content_type": "application/yaml" |
|
} |
|
], |
|
"conversation_id": "123e4567-e89b-12d3-a456-426614174000", |
|
"model": "gpt-3.5-turbo", |
|
"provider": "openai", |
|
"query": "write a deployment yaml for the mongodb image" |
|
} |
|
] |
|
}, |
Another example can be found on PDL(prompt declaration language) project:
https://github.com/IBM/prompt-declaration-language/blob/572373a09e2d105cf6712859d4be5fb371ba1051/examples/tutorial/calling_llm_with_input_messages.pdl#L5-L10
This, as far as I know:
- Will not break any backward compatibility, it's a new parameter.
- It might override the system prompt given in the olsconfig.yaml.
- Can break the history message placeholder, it depends on where it's located.
Steps needed
Questions:
- What I should do to get it merged?
- Is this a valid approach?
Type: feature-request
Description:
When using complex scenarios, we need to teach the LLM how to generate the
response. This technique is a few-shot prompting, where a few messages are part
of the prompt before the human question.
Ollama implemented this using the message array, where the end user define it:
https://github.com/ollama/ollama/blob/dc6fe820512d1046f3a342e384baa64b8ce1758c/docs/api.md?plain=1#L451-L457
I think that in this case, it'll be cool to use something similar which can be appended here:
service/docs/openapi.json
Lines 509 to 595 in 5360e36
Another example can be found on PDL(prompt declaration language) project:
https://github.com/IBM/prompt-declaration-language/blob/572373a09e2d105cf6712859d4be5fb371ba1051/examples/tutorial/calling_llm_with_input_messages.pdl#L5-L10
This, as far as I know:
Steps needed
Questions: