Skip to main content

internal/infra/inference/prompt.go

internal/infra/inference · 185 lines · 10 declarations · source

Declarations​

var extractionPromptFile​

var extractionPromptFile []byte

The extraction prompt is a file because it is prose, and prose is reviewed as prose. It is compiled in because it decides what this system extracts, and extraction feeds the write path: a prompt the filesystem can change is a path by which whoever can write next to the binary changes what ends up in a customer's memory. Embedding keeps the review benefit and closes that.

source

const relationsPlaceholder​

const relationsPlaceholder = "{{relations}}"

relationsPlaceholder is where the vocabulary is rendered.

The predicate table is not written in the file: it is enforced by a foreign key, and a second copy of a closed set is a second place for it to drift. The file marks the position, the code renders the rows.

source

type promptFile​

type promptFile struct {
	ModelContract struct {
		Temperature    float32 `yaml:"temperature"`
		ResponseFormat string  `yaml:"response_format"`
	} `yaml:"model_contract"`

	Task         string `yaml:"task"`
	Relations    string `yaml:"relations"`
	QuoteRule    string `yaml:"quote_rule"`
	Polarity     string `yaml:"polarity"`
	Tense        string `yaml:"tense"`
	ReplyShape   string `yaml:"reply_shape"`
	DataBoundary string `yaml:"data_boundary"`
}

promptFile is the wording, and only the wording.

Every field is required. A section that may be absent is a prompt that can be silently shortened by an edit, and the section most worth removing — the one telling the model the message is data — is the one whose absence produces no error and no visible change until somebody plants a fact.

source

var extractionPrompt​

var extractionPrompt = mustLoadPrompt(extractionPromptFile)

extractionPrompt is parsed once, at startup.

A panic is right here. The file is embedded, so a failure is not a runtime condition that might clear — it is a binary that must not start, and it is caught by any test run before it is ever caught by a deployment.

source

func mustLoadPrompt​

func mustLoadPrompt(raw []byte) promptFile

source

func loadPrompt​

func loadPrompt(raw []byte) (promptFile, error)

loadPrompt parses the prompt file and refuses one that could not compose a whole prompt.

source

func systemPrompt​

func systemPrompt(vocabulary domain.Ontology) string

systemPrompt composes the prompt: the file's wording, in an order the file does not choose.

The order is fixed here, and data_boundary is last, because that section's job is to be the final word before somebody else's text arrives. A file that could reorder the sections could put the injection boundary in the middle, where an instruction after it appears to supersede it.

source

func renderVocabulary​

func renderVocabulary(vocabulary domain.Ontology) string

renderVocabulary writes the relations out in full, one per line.

Each relation carries its description and the kind of thing its object should be. Naming the relations without saying what they mean produces confident misuse — `part_of` for employment, `has_status` for anything — and a wrong relation is worse than an unmapped one, because it is admitted and nothing downstream can tell it apart from a right one.

The order is the vocabulary's own, which is the order the database holds it in. Stable, so that two extractions of one message are given the same list: a prompt that reorders produces different output for the same input, and extraction quality stops being measurable.

source

type reportPromptFileShape​

type reportPromptFileShape struct {
	ModelContract struct {
		Temperature    float32 `yaml:"temperature"`
		ResponseFormat string  `yaml:"response_format"`
	} `yaml:"model_contract"`

	Task         string `yaml:"task"`
	Grounding    string `yaml:"grounding"`
	Shape        string `yaml:"shape"`
	ReplyShape   string `yaml:"reply_shape"`
	DataBoundary string `yaml:"data_boundary"`
}

reportPromptFileShape is the report prompt's wording, and only the wording.

Every field is required, for the reason the extraction prompt's are: a section that may be absent is a prompt that can be silently shortened by an edit, and the section most worth removing — the one telling the model the material is data — is the one whose absence produces no error and no visible change until somebody plants an instruction in a quote.

source

func loadReportPrompt​

func loadReportPrompt(raw []byte) (reportPromptFileShape, error)

loadReportPrompt parses the report prompt and refuses one that could not compose a whole prompt.

source