rony-chat-bot/internal/config/config.go

package config

import (
	"fmt"
	"os"
	"strings"

	"gopkg.in/yaml.v3"
)

type Server struct {
	Host           string   `yaml:"host"`
	Port           int      `yaml:"port"`
	ReadTimeoutMS  int      `yaml:"read_timeout_ms"`
	CORSOrigins    []string `yaml:"cors_origins"`
	RateLimit      RateLimit `yaml:"rate_limit"`
}

type RateLimit struct {
	RequestsPerMinute int `yaml:"requests_per_minute"`
	Burst             int `yaml:"burst"`
}

// Provider is one upstream LLM. The sampling fields (Temperature through
// PresencePenalty) are forwarded verbatim to the llama.cpp adapter; a zero
// value means "don't send it", so llama-server's own default applies. They
// matter more than they look: small instruct models ship with vendor-
// recommended sampling (Gemma 3 wants temp 1.0 / top_k 64 / top_p 0.95) and
// drifting off it makes them terse and repetitive.
type Provider struct {
	Name        string  `yaml:"name"`
	Type        string  `yaml:"type"`
	Model       string  `yaml:"model,omitempty"`
	Endpoint    string  `yaml:"endpoint,omitempty"`
	ContextSize int     `yaml:"context_size,omitempty"`
	MaxTokens   int     `yaml:"max_tokens,omitempty"`
	Temperature float32 `yaml:"temperature,omitempty"`
	TopK        int     `yaml:"top_k,omitempty"`
	TopP        float32 `yaml:"top_p,omitempty"`
	MinP        float32 `yaml:"min_p,omitempty"`
	// RepeatPenalty maps to llama.cpp's repeat_penalty. Leave unset (0) for
	// Gemma 3 — Google's recommended config is no repetition penalty at all,
	// and a value >1 visibly degrades its prose.
	RepeatPenalty   float32 `yaml:"repeat_penalty,omitempty"`
	PresencePenalty float32 `yaml:"presence_penalty,omitempty"`
	APIKeyEnv       string  `yaml:"api_key_env,omitempty"`
	Default         bool    `yaml:"default,omitempty"`
}

type RAG struct {
	Enabled  bool   `yaml:"enabled"`
	DataPath string `yaml:"data_path"`
	// DocsPath is an optional second directory of markdown that is indexed
	// and searchable but is NOT part of the project catalogue: a CV, an about
	// page, a FAQ. Without it the only way to make the CV retrievable is to
	// drop it in data_path, where it then gets announced as one of Victor's
	// projects. Both directories accept .md and .mdx.
	DocsPath string `yaml:"docs_path"`
	DBPath   string `yaml:"db_path"`
	TopK     int    `yaml:"top_k"`
	Tokenize string `yaml:"tokenize"`
	// IncludeCatalog injects the full list of indexed projects into the
	// system prompt on every turn. Costs a few tokens per project and stops
	// small models from inventing project names when asked to enumerate —
	// top-K retrieval can't answer "list everything" by construction.
	IncludeCatalog bool `yaml:"include_catalog"`
}

// Embeddings configures the semantic half of retrieval. Disabled by default,
// in which case the bot uses keyword search only.
//
// Run the endpoint with:
//
//	llama-server -m nomic-embed-v2-moe.Q5_K_M.gguf --port 9200 \
//	  --embedding --pooling mean --ctx-size 2048 --parallel 1 \
//	  --device none --threads 2
//
// Measured at 0.61 GB RSS on CPU. It has to stay resident: the corpus is
// embedded once at index time, but every visitor question must be embedded
// before it can be compared.
type Embeddings struct {
	Enabled   bool   `yaml:"enabled"`
	Endpoint  string `yaml:"endpoint"`
	Model     string `yaml:"model,omitempty"`
	BatchSize int    `yaml:"batch_size,omitempty"`
	TimeoutMS int    `yaml:"timeout_ms,omitempty"`
}

type Persona struct {
	Name     string `yaml:"name"`
	Tone     string `yaml:"tone"`
	Language string `yaml:"language"`
}

type Logging struct {
	Level  string `yaml:"level"`
	Format string `yaml:"format"`
	Output string `yaml:"output"`
}

// Compaction controls automatic context-window management. When the most
// recent request's input tokens exceed threshold_ratio × the provider's
// reported MaxContextWindow, the older portion of the conversation is
// summarized by the LLM and replaced with a single system-role note.
// Recent turns (keep_recent_turns) are always preserved verbatim.
//
// Disabled by default; flip enabled: true in the YAML to opt in. The
// model's MaxContextWindow comes from llm.LLMClient.Capabilities(), so
// the same setting works for llamacpp, ollama, openai and anthropic
// without per-provider configuration.
type Compaction struct {
	Enabled     bool    `yaml:"enabled"`
	// ThresholdRatio is the fraction of the context window that, once
	// exceeded by the previous turn's input tokens, triggers compaction.
	// Range 0–1; default 0.75 (75% of the window).
	ThresholdRatio float64 `yaml:"threshold_ratio"`
	// KeepRecentTurns is how many of the most recent user turns (each
	// being a user message plus everything up to the next user message)
	// to keep verbatim after compaction. Older turns are folded into the
	// summary. Default 4.
	KeepRecentTurns int `yaml:"keep_recent_turns"`
	// SummarySystemPrompt is the instruction sent to the LLM when it
	// summarizes the older turns. Empty falls back to a built-in
	// bilingual prompt that auto-matches the conversation language.
	SummarySystemPrompt string `yaml:"summary_system_prompt"`
}

type Config struct {
	Server       Server     `yaml:"server"`
	Providers    []Provider `yaml:"providers"`
	RAG          RAG        `yaml:"rag"`
	Embeddings   Embeddings `yaml:"embeddings"`
	Persona      Persona    `yaml:"persona"`
	SystemPrompt string     `yaml:"system_prompt"`
	// SystemPromptES is the Spanish rendition of SystemPrompt, used when the
	// visitor writes in Spanish. Optional; empty falls back to SystemPrompt.
	//
	// This is not a convenience — it's the only thing that reliably keeps a
	// small model answering in Spanish. An instruction like "reply in the
	// user's language" buried in an otherwise English prompt loses to the
	// sheer mass of English around it: measured on gemma-3-1b, 1 of 5 Spanish
	// questions came back in Spanish. Translating the prompt itself took that
	// to 4 of 5. (Few-shot Spanish examples also score well, but a 1B model
	// copies them verbatim instead of answering — see the note on
	// `system_prompt`.)
	SystemPromptES string `yaml:"system_prompt_es"`
	Compaction   Compaction `yaml:"compaction"`
	Logging      Logging    `yaml:"logging"`
}

func Load(path string) (*Config, error) {
	raw, err := os.ReadFile(path)
	if err != nil {
		return nil, fmt.Errorf("read config %s: %w", path, err)
	}

	expanded := os.ExpandEnv(string(raw))

	var cfg Config
	if err := yaml.Unmarshal([]byte(expanded), &cfg); err != nil {
		return nil, fmt.Errorf("parse config %s: %w", path, err)
	}

	if err := cfg.validate(); err != nil {
		return nil, err
	}
	return &cfg, nil
}

func (c *Config) validate() error {
	if len(c.Providers) == 0 {
		return fmt.Errorf("config: at least one provider must be configured")
	}
	defaultCount := 0
	for _, p := range c.Providers {
		if p.Default {
			defaultCount++
		}
	}
	if defaultCount == 0 {
		c.Providers[0].Default = true
	} else if defaultCount > 1 {
		return fmt.Errorf("config: multiple providers marked as default")
	}
	if c.Server.Port == 0 {
		c.Server.Port = 7331
	}
	if c.Server.ReadTimeoutMS == 0 {
		c.Server.ReadTimeoutMS = 30000
	}
	if c.RAG.TopK == 0 {
		c.RAG.TopK = 5
	}
	if c.RAG.Tokenize == "" {
		c.RAG.Tokenize = "unicode61"
	}
	if c.Compaction.Enabled {
		if c.Compaction.ThresholdRatio <= 0 || c.Compaction.ThresholdRatio > 1 {
			c.Compaction.ThresholdRatio = 0.75
		}
		if c.Compaction.KeepRecentTurns <= 0 {
			c.Compaction.KeepRecentTurns = 4
		}
	}
	return nil
}

func (c *Config) DefaultProvider() *Provider {
	for i := range c.Providers {
		if c.Providers[i].Default {
			return &c.Providers[i]
		}
	}
	return &c.Providers[0]
}

func (c *Config) Addr() string {
	return fmt.Sprintf("%s:%d", c.Server.Host, c.Server.Port)
}

func splitPath(path string) []string {
	return strings.Split(path, string(os.PathSeparator))
}