config

Documentation

Overview

Package config loads and validates the andino-kb YAML configuration.

Decoding is strict: unknown fields are errors, so a typo in a pipeline definition fails at startup instead of silently indexing nothing. ${VAR} references are expanded from the environment before parsing; unset variables expand to the empty string.

Index

• Constants
• type APIKey
• type Backend
• type ChatModel
• type Chunking
• type Config
  • func Load(path string) (*Config, error)
  • func (c *Config) ChatModelByName(name string) (ChatModel, Backend, error)
  • func (c *Config) EmbeddingModelFor(kb *KnowledgeBase) (EmbeddingModel, Backend, error)
  • func (c *Config) Validate() error
• type Contextual
• type Defaults
• type EmbeddingModel
• type Inference
• type KnowledgeBase
• type OCR
• type RerankModel
• type Server
• type Source
• type Storage

Constants

const ManagedSourceName = "managed"

ManagedSourceName is the implicit source name for agent-written documents in writable knowledge bases.

Types

type APIKey

type APIKey struct {
	Key   string `yaml:"key"`
	Scope string `yaml:"scope"`
}

APIKey grants access to the REST and MCP APIs. Scope "read" allows search and reads; "readwrite" additionally allows store/delete on writable KBs.

type Backend

type Backend struct {
	Name    string `yaml:"name"`
	BaseURL string `yaml:"base_url"`
	APIKey  string `yaml:"api_key"`
}

type ChatModel

type ChatModel struct {
	Name      string `yaml:"name"`
	Backend   string `yaml:"backend"`
	Model     string `yaml:"model"`
	MaxTokens int    `yaml:"max_tokens"`
	// ExtraBody is merged into the /v1/chat/completions request body.
	// Needed e.g. to disable reasoning on thinking-first models
	// (llama.cpp/vLLM: chat_template_kwargs: {enable_thinking: false}),
	// whose reasoning otherwise consumes max_tokens and returns empty
	// content.
	ExtraBody map[string]any `yaml:"extra_body"`
}

ChatModel is a chat-completions model used for index-time work such as contextual retrieval.

type Chunking

type Chunking struct {
	Strategy      string `yaml:"strategy"`
	MaxTokens     int    `yaml:"max_tokens"`
	OverlapTokens int    `yaml:"overlap_tokens"`
}

type Config

type Config struct {
	Server         Server          `yaml:"server"`
	Storage        Storage         `yaml:"storage"`
	Inference      Inference       `yaml:"inference"`
	Defaults       Defaults        `yaml:"defaults"`
	KnowledgeBases []KnowledgeBase `yaml:"knowledge_bases"`
}

func Load

func Load(path string) (*Config, error)

Load reads, expands, parses and validates a config file.

func (*Config) ChatModelByName

func (c *Config) ChatModelByName(name string) (ChatModel, Backend, error)

ChatModelByName resolves a chat model reference to its definition and backend.

func (*Config) EmbeddingModelFor

func (c *Config) EmbeddingModelFor(kb *KnowledgeBase) (EmbeddingModel, Backend, error)

EmbeddingModelFor resolves a KB's embedding model definition.

func (*Config) Validate

func (c *Config) Validate() error

type Contextual

type Contextual struct {
	Enabled bool   `yaml:"enabled"`
	Model   string `yaml:"model"` // ref into inference.chat_models
}

Contextual enables contextual retrieval for a knowledge base: an LLM generates a short situating context per chunk at index time, which is embedded and BM25-indexed alongside the text.

type Defaults

type Defaults struct {
	Chunking       Chunking `yaml:"chunking"`
	EmbeddingModel string   `yaml:"embedding_model"`
	RerankModel    string   `yaml:"rerank_model"`
	OCR            *OCR     `yaml:"ocr"`
}

type EmbeddingModel

type EmbeddingModel struct {
	Name       string `yaml:"name"`
	Backend    string `yaml:"backend"`
	Model      string `yaml:"model"`
	Dimensions int    `yaml:"dimensions"`
	BatchSize  int    `yaml:"batch_size"`
	MaxRetries int    `yaml:"max_retries"`
}

type Inference

type Inference struct {
	Backends        []Backend        `yaml:"backends"`
	EmbeddingModels []EmbeddingModel `yaml:"embedding_models"`
	RerankModels    []RerankModel    `yaml:"rerank_models"`
	ChatModels      []ChatModel      `yaml:"chat_models"`
}

type KnowledgeBase

type KnowledgeBase struct {
	Name           string    `yaml:"name"`
	Description    string    `yaml:"description"`
	Writable       bool      `yaml:"writable"`
	Sources        []Source  `yaml:"sources"`
	Chunking       *Chunking `yaml:"chunking"`
	EmbeddingModel string    `yaml:"embedding_model"`
	RerankModel    string    `yaml:"rerank_model"`
	// RerankDefault decides whether searches rerank when no per-request
	// override is given: "on" (default) or "off". With "off" the reranker
	// stays available to requests that ask for rerank: true.
	RerankDefault string      `yaml:"rerank_default"`
	Contextual    *Contextual `yaml:"contextual"`
	OCR           *OCR        `yaml:"ocr"`
}

type OCR

type OCR struct {
	Enabled bool   `yaml:"enabled"`
	Model   string `yaml:"model"` // ref into inference.chat_models (must be vision-capable)
}

OCR enables transcription of scanned PDF pages through a vision-capable chat model at index time.

type RerankModel

type RerankModel struct {
	Name    string `yaml:"name"`
	Backend string `yaml:"backend"`
	Model   string `yaml:"model"`
}

type Server

type Server struct {
	Bind      string   `yaml:"bind"`
	DataDir   string   `yaml:"data_dir"`
	APIKeys   []APIKey `yaml:"api_keys"`
	LogLevel  string   `yaml:"log_level"`
	LogFormat string   `yaml:"log_format"`
}

type Source

type Source struct {
	Name string `yaml:"name"`
	Type string `yaml:"type"` // localdir | git | s3

	// localdir
	Path       string   `yaml:"path"`
	Include    []string `yaml:"include"`
	Exclude    []string `yaml:"exclude"`
	Watch      bool     `yaml:"watch"`
	DebounceMS int      `yaml:"debounce_ms"`

	// git
	URL          string        `yaml:"url"`
	Branch       string        `yaml:"branch"`
	Paths        []string      `yaml:"paths"` // also used by s3
	PollInterval time.Duration `yaml:"poll_interval"`
	TokenEnv     string        `yaml:"token_env"`

	// s3
	Bucket    string `yaml:"bucket"`
	Prefix    string `yaml:"prefix"`
	Region    string `yaml:"region"`
	Endpoint  string `yaml:"endpoint"`   // custom endpoint for MinIO/compatible
	PathStyle bool   `yaml:"path_style"` // path-style addressing (MinIO)
}

Source is a single ingestion pipeline. Type-specific fields are flat; the validator enforces which apply to which type.

type Storage

type Storage struct {
	Provider string         `yaml:"provider"`
	Options  map[string]any `yaml:"options"`
}