omnia-consolidated/packages/llm/README.md

@omnia/llm

LLM abstraction layer providing pluggable, database-backed provider instances for generative and embedding tasks.

Architecture Overview

The system is built around three layers:

1. Interfaces — contracts that all providers implement
2. Provider Manager — SQLite-backed CRUD for persisted provider instances
3. Provider Resolver — runtime instantiation of concrete provider classes from stored instances

graph TD
    subgraph Interfaces
        ILP["ILLMProvider"]
        IEP["IEmbeddingProvider"]
        MPI["ModelProviderInstance"]
    end

    subgraph Concrete Providers
        GP["GeminiProvider"]
        ORP["OpenRouterProvider"]
        MP["MockLLMProvider"]
        GEP["GeminiEmbeddingProvider"]
        MEP["MockEmbeddingProvider"]
    end

    subgraph Storage
        PM["ProviderManager"]
        DB[("settings.db\nprovider_instances")]
        DBMAP[("settings.db\nprovider_mappings")]
    end

    subgraph Resolution
        PR["resolveProviders()"]
    end

    GP -->|implements| ILP
    ORP -->|implements| ILP
    MP -->|implements| ILP
    GEP -->|implements| IEP
    MEP -->|implements| IEP

    PM -->|reads/writes| DB
    PM -->|reads/writes| DBMAP
    PM -->|returns| MPI

    PR -->|queries| PM
    PR -->|instantiates| GP
    PR -->|instantiates| ORP
    PR -->|instantiates| GEP
    PR -->|fallback| MP
    PR -->|fallback| MEP

Core Interfaces

Defined in llm.ts:

ILLMProvider

The primary contract for generative (text-to-structured-data) providers.

Member	Type	Description
providerName	string	Human-readable provider label
maxContext	number?	Maximum context window in tokens
generateStructuredResponse<T>(request)	Promise<LLMResponse<T>>	Sends a prompt + Zod schema → returns parsed, typed output
lastCalls	LLMCallRecord[]?	Audit trail of recent calls (prompts + usage)

IEmbeddingProvider

Contract for text-to-vector embedding providers.

Member	Type	Description
providerName	string	Human-readable provider label
embed(text)	Promise<number[]>	Returns a dense vector embedding of the input text

LLMRequest<T>

Input to generateStructuredResponse:

{
  systemPrompt: string;    // System-level instructions
  userContext:   string;    // User/task-specific context
  schema:        T;         // Zod schema — output is validated against this
  temperature?:  number;    // Sampling temperature (optional)
}

LLMResponse<T>

Output from generateStructuredResponse:

{
  success: boolean;
  data?:   T;               // Parsed, schema-validated output
  error?:  string;          // Error message on failure
  usage?:  {
    inputTokens:          number;
    outputTokens:         number;
    totalTokens:          number;
    modelName?:           string;
    providerInstanceName?: string;
    maxContext?:           number;
  };
}

ModelProviderInstance

The persisted configuration record for a single provider instance:

{
  id:           string;                     // Unique ID ("provider-<timestamp>")
  name:         string;                     // User-facing name ("Gemini (Env)")
  providerName: string;                     // Provider type key ("google-genai" | "openrouter" | "mock")
  apiKey:       string;                     // API key
  isActive:     boolean;                    // Whether this is the active instance for its type
  modelName?:   string;                     // Specific model to use
  type:         "generative" | "embedding"; // Instance category
  maxContext?:  number;                     // Context window limit
}

ModelProviderMeta

Static metadata for each available provider type (used by the UI's provider picker):

Member	Type	Description
id	string	"google-genai" | "openrouter" | "ollama" | "mock"
displayName	string	Human-readable name
description	string	Human-readable description
defaultModel	string	Default generative model
defaultEmbeddingModel	string	Default embedding model

The AVAILABLE_PROVIDERS constant exports all four provider metas.

Provider Manager

ProviderManager is a static class that provides full CRUD over provider instances, backed by a SQLite database (data/settings.db at the workspace root).

Storage

The database is auto-created on first access. The table schema:

CREATE TABLE IF NOT EXISTS provider_instances (
  id           TEXT PRIMARY KEY,
  name         TEXT NOT NULL,
  providerName TEXT NOT NULL,
  apiKey       TEXT NOT NULL,
  isActive     INTEGER NOT NULL DEFAULT 0,
  modelName    TEXT,
  type         TEXT NOT NULL DEFAULT 'generative',
  maxContext   INTEGER
);

A second table stores per-task provider overrides:

CREATE TABLE IF NOT EXISTS provider_mappings (
  task                TEXT PRIMARY KEY,
  providerInstanceId  TEXT NOT NULL
);

API

Method	Signature	Description
list()	→ ModelProviderInstance[]	Returns all saved instances
create(name, providerName, apiKey, modelName?, type?, maxContext?)	→ ModelProviderInstance	Creates a new instance. Auto-activates if it's the first of its type
delete(id)	→ void	Removes an instance. If it was active, auto-promotes the next instance of the same type
setActive(id)	→ void	Deactivates all instances of the same type, then activates the target
update(id, name, providerName, apiKey?, modelName?, type?, maxContext?)	→ void	Updates an existing instance. If apiKey is empty/omitted, the existing key is preserved
getActive(type?)	→ ModelProviderInstance | null	Returns the currently active instance for the given type ("generative" by default)
getMappings()	→ Record<string, string>	Returns all task → providerInstanceId mappings
setMapping(task, providerInstanceId)	→ void	Sets or removes (if providerInstanceId is empty) a task-specific mapping

Active Instance Invariants

• Only one active instance per type — setActive() deactivates all sibling instances before activating the target.
• Auto-promotion on delete — if the deleted instance was active, the first remaining instance of the same type is promoted.
• Auto-activation on create — if no active instance exists for the type, the new instance is automatically activated.

Environment Variable Bootstrap

On first database access (and if the provider_instances table is empty), the manager auto-seeds instances from environment variables:

flowchart TD
    A["getSettingsDb() called"] --> B{"DB has 0 rows?"}
    B -- No --> Z["Return DB"]
    B -- Yes --> C{"GOOGLE_API_KEY set?"}
    C -- Yes --> D["Insert 'Gemini (Env)'\ntype: generative, active: true"]
    D --> E["Insert 'Gemini Embed (Env)'\ntype: embedding, active: true"]
    E --> F{"OPENROUTER_API_KEY set?"}
    C -- No --> F
    F -- Yes --> G["Insert 'OpenRouter (Env)'\ntype: generative\nactive: only if no Google key"]
    F -- No --> Z
    G --> Z

This same bootstrap logic is duplicated inside getActive() as a safety net — if the DB is empty at query time, it re-attempts the same env-var seeding.

Fallback Chain in getActive()

When no active row is found for the requested type:

1. DB query for isActive=1 AND type=<requested>
   ├── Found → return it
   └── Not found
       ├── DB is empty → bootstrap from env vars → retry query
       │   ├── Found → return it
       │   └── Still empty → promote first row of same type
       │       ├── Found → activate & return
       │       └── None → return null
       └── DB has rows but none active for this type
           → promote first row of same type (same as above)

2. On any DB error (catch block) → direct env var fallback
   ├── GOOGLE_API_KEY → synthetic "Gemini (Env Fallback)" instance
   ├── OPENROUTER_API_KEY → synthetic "OpenRouter (Env Fallback)" instance
   └── Neither → return null

Available Providers

Google Gemini — GeminiProvider

Property	Value
File	providers/google-genai.ts
Provider ID	google-genai
SDK	@langchain/google-genai (ChatGoogleGenerativeAI)
Default Model	gemini-2.5-flash
Default Embedding Model	gemini-embedding-001
Default Max Context	32768
Type	Generative

Key resolution in the constructor follows this cascade:

1. Explicit apiKey argument       → use it
2. ProviderManager.getActive()    → if providerName matches "google-genai"
3. GOOGLE_API_KEY env var         → final fallback
4. None found                     → throw Error

Also exports GeminiEmbeddingProvider (implements IEmbeddingProvider) using the same key resolution pattern but querying for the "embedding" type.

OpenRouter — OpenRouterProvider

Property	Value
File	providers/openrouter.ts
Provider ID	openrouter
SDK	@langchain/openrouter (ChatOpenRouter)
Default Model	google/gemini-2.5-flash
Default Embedding Model	openai/text-embedding-3-small
Default Max Context	32768
Type	Generative only (no embedding provider)

Same three-step key resolution as Gemini (explicit → ProviderManager → env var), using OPENROUTER_API_KEY.

Ollama — OllamaProvider

Property	Value
File	providers/ollama.ts
Provider ID	ollama
SDK	@langchain/ollama (ChatOllama)
Default Model	llama3.1
Default Embedding Model	nomic-embed-text
Default Max Context	32768
Type	Generative + Embedding

Ollama runs locally — no API key is required. The apiKey field in ModelProviderInstance is repurposed to store the Ollama server base URL (default: http://localhost:11434).

Key resolution in the constructor:

1. Explicit baseUrl argument        → use it
2. ProviderManager.getActive()      → if providerName matches "ollama"
                                      (apiKey field = base URL)
3. Default                          → http://localhost:11434

Also exports OllamaEmbeddingProvider (implements IEmbeddingProvider), which uses the same resolution pattern against the "embedding" type instance. The default embedding model is nomic-embed-text.

Tip

To get started: ollama pull llama3.1 and ollama pull nomic-embed-text. Then create a provider instance with apiKey = http://localhost:11434.

Mock — MockLLMProvider

Property	Value
File	providers/mock.ts
Provider ID	mock
Type	Generative + Embedding

Stateless mock for testing and offline development:

• Generative (MockLLMProvider): Takes an array of canned responses at construction. Returns them in order, one per call. Returns { success: false, error: "Mock responses exhausted" } when depleted.
• Embedding (MockEmbeddingProvider): Returns a deterministic 768-dimensional vector derived from the input text using Math.sin.

Provider Resolution (Runtime)

The resolveProviders() function (in apps/gui) instantiates all providers needed for a simulation session. It resolves six provider slots:

Slot	Type	Task Key
actorProvider	Generative	"actor-prose"
validatorProvider	Generative	"llm-validator"
decoderProvider	Generative	"intent-decoder"
timedeltaProvider	Generative	"timedelta"
handoffProvider	Generative	"handoff"
embeddingProvider	Embedding	"embeddings"

Generative Resolution Order

For each generative slot (resolveGenerative(task)):

1. Task-specific mapping     → mappings[task] → find instance by ID
2. Active generative instance → ProviderManager.getActive("generative")
3. Fallback instance         → options.fallbackInstance (if provided)
4. GOOGLE_API_KEY env var    → auto-create via ProviderManager.create()
5. No provider available     → throw Error (if required) or MockLLMProvider

Embedding Resolution Order

For the embedding slot (resolveEmbedding()):

1. Task-specific mapping     → mappings["embeddings"] → find instance by ID
2. Active embedding instance → ProviderManager.getActive("embedding")
3. GOOGLE_API_KEY env var    → auto-create via ProviderManager.create()
4. No provider available     → throw Error (if required) or MockEmbeddingProvider

Instance → Class Mapping

The buildLLMProvider() and buildEmbeddingProvider() functions perform the final dispatch:

providerName	Generative Class	Embedding Class
"google-genai"	GeminiProvider	GeminiEmbeddingProvider
"openrouter"	OpenRouterProvider	(falls through to mock)
"ollama"	OllamaProvider	OllamaEmbeddingProvider
(anything else)	MockLLMProvider	MockEmbeddingProvider

Structured Output

All real providers use LangChain's .withStructuredOutput(schema, { includeRaw: true }) pattern:

const structuredModel = this.model.withStructuredOutput(request.schema, {
  includeRaw: true,
});
const result = await structuredModel.invoke([
  { role: "system", content: request.systemPrompt },
  { role: "user", content: request.userContext },
]);

This sends the Zod schema to the model as a structured output constraint. The response includes both parsed (schema-validated data) and raw (full API response with usage metadata).

Configuration

config.ts parses environment variables using Zod:

Variable	Required	Description
GOOGLE_API_KEY	No	Google Gemini API key
OPENROUTER_API_KEY	No	OpenRouter API key

Both are optional because providers can also be configured through the database via the GUI settings page.

File Map

packages/llm/
├── src/
│   ├── index.ts              # Re-exports everything
│   ├── llm.ts                # Interfaces, types, AVAILABLE_PROVIDERS
│   ├── config.ts             # Env var parsing (Zod)
│   ├── provider-manager.ts   # ProviderManager (SQLite CRUD)
│   └── providers/
│       ├── google-genai.ts   # GeminiProvider + GeminiEmbeddingProvider
│       ├── ollama.ts         # OllamaProvider + OllamaEmbeddingProvider
│       ├── openrouter.ts     # OpenRouterProvider
│       └── mock.ts           # MockLLMProvider + MockEmbeddingProvider
├── tests/
│   ├── mock.test.ts
│   ├── openrouter.test.ts
│   └── provider-manager.test.ts
└── package.json