llm

type Adapter ¶

type Adapter interface {
	// Classify returns category, tags, sensitivity, and a one-line summary.
	// Called when a note is added without explicit metadata.
	Classify(ctx context.Context, content string, categories []model.Category, tags []string) (ClassifyResult, error)

	// Answer performs RAG generation: query + retrieved notes → structured response.
	// The response carries both the natural-language answer and a flag telling the
	// caller whether the notes were actually useful (so callers can hide "Sources"
	// blocks when the LLM rejected every candidate).
	Answer(ctx context.Context, query string, notes []*model.Note) (AnswerResult, error)

	// Summarize returns a one-line description of the note content.
	Summarize(ctx context.Context, content string) (string, error)
}

Adapter defines the contract for all LLM backends. Implementations: claude.go, ollama.go.

type AnswerResult ¶

type AnswerResult struct {
	Answer      string
	Matched     bool
	UsedNoteIDs []string
}

AnswerResult is the structured output from RAG answer generation.

Matched is true iff the LLM concluded that at least one of the provided notes contained the answer. UsedNoteIDs holds short-prefix IDs of the notes the LLM actually used; the search layer filters the retrieved notes down to these so the CLI only displays sources that contributed.

When Matched is false, UsedNoteIDs is empty and Answer is the LLM's natural-language explanation that the answer is not in the notes.

type ClassifyResult ¶

type ClassifyResult struct {
	Category    string
	Tags        []string
	IsSensitive bool
	Summary     string
}

ClassifyResult is the structured output from LLM classification.

type ClaudeAdapter ¶

type ClaudeAdapter struct {
	// contains filtered or unexported fields
}

ClaudeAdapter implements LLMAdapter using the Anthropic Claude API.

func NewClaudeAdapter(apiKey, modelName string) *ClaudeAdapter

func (c *ClaudeAdapter) Answer(ctx context.Context, query string, notes []*model.Note) (AnswerResult, error)

func (c *ClaudeAdapter) Classify(ctx context.Context, content string, categories []model.Category, tags []string) (ClassifyResult, error)

func (c *ClaudeAdapter) Summarize(ctx context.Context, content string) (string, error)

type OllamaAdapter ¶

type OllamaAdapter struct {
	// contains filtered or unexported fields
}

OllamaAdapter implements LLMAdapter using a local Ollama server.

func NewOllamaAdapter(baseURL, modelName string) *OllamaAdapter

func (o *OllamaAdapter) Answer(ctx context.Context, query string, notes []*model.Note) (AnswerResult, error)

func (o *OllamaAdapter) Classify(ctx context.Context, content string, categories []model.Category, tags []string) (ClassifyResult, error)

func (o *OllamaAdapter) Summarize(ctx context.Context, content string) (string, error)

type OpenAIAdapter ¶

type OpenAIAdapter struct {
	// contains filtered or unexported fields
}

OpenAIAdapter implements the LLM Adapter using the OpenAI Chat Completions API. It talks plain HTTP (no SDK dependency) so the binary stays lean and `go install` needs no extra modules, mirroring the Ollama adapter's style, but pointed at OpenAI with Bearer auth and JSON-mode responses.

func NewOpenAIAdapter(apiKey, modelName, baseURL string) *OpenAIAdapter

func (o *OpenAIAdapter) Answer(ctx context.Context, query string, notes []*model.Note) (AnswerResult, error)

func (o *OpenAIAdapter) Classify(ctx context.Context, content string, categories []model.Category, tags []string) (ClassifyResult, error)

func (o *OpenAIAdapter) Summarize(ctx context.Context, content string) (string, error)