graphile-llm 0.7.2 → 0.8.0

package/esm/config-cache.js

@@ -0,0 +1,143 @@
+/**
+ * config-cache — Per-database LLM billing configuration cache
+ *
+ * Caches resolved billing function names per database_id.
+ * Uses an LRU cache with TTL so config changes propagate within a bounded window
+ * without requiring a server restart.
+ *
+ * Resolution flow:
+ *   Billing config from `metaschema_modules_public.billing_module`
+ *   (schema name + function names for record_usage, check_billing_quota)
+ *
+ * All queries run through the Graphile `withPgClient` callback, which gives us
+ * a client connected to the tenant database with proper role settings.
+ *
+ * The LLM module config (provider, model, etc.) is already resolved by the
+ * LlmModulePlugin at schema-build time. This cache handles the runtime-only
+ * billing piece.
+ */
+import { ModuleConfigCache } from 'graphile-cache';
+// ─── SQL Queries ────────────────────────────────────────────────────────────
+/**
+ * Check if the billing_module table exists before querying it.
+ * This prevents hard errors on databases that don't have the billing
+ * module provisioned (the metaschema_modules_public schema or the
+ * billing_module table might not exist at all).
+ */
+const BILLING_MODULE_SQL = `
+  SELECT
+    s.schema_name AS public_schema,
+    ps.schema_name AS private_schema,
+    bm.record_usage_function
+  FROM metaschema_modules_public.billing_module bm
+  JOIN metaschema_public.schema s ON bm.schema_id = s.id
+  JOIN metaschema_public.schema ps ON bm.private_schema_id = ps.id
+  WHERE bm.database_id = $1
+  LIMIT 1
+`;
+/**
+ * Resolve the inference log module's schema and table name.
+ */
+const INFERENCE_LOG_MODULE_SQL = `
+  SELECT
+    s.schema_name AS schema,
+    ilm.inference_log_table_name AS table_name
+  FROM metaschema_modules_public.inference_log_module ilm
+  JOIN metaschema_public.schema s ON ilm.schema_id = s.id
+  WHERE ilm.database_id = $1
+  LIMIT 1
+`;
+// ─── Cache ──────────────────────────────────────────────────────────────────
+const billingCache = new ModuleConfigCache({
+    name: 'billing-config',
+    ttlMs: 5 * 60 * 1000, // 5 minutes
+    max: 50,
+});
+// ─── Resolution Functions ───────────────────────────────────────────────────
+/**
+ * SQL to check if a schema exists. Used as a guard before querying
+ * metaschema tables that may not be provisioned.
+ */
+const SCHEMA_EXISTS_SQL = `
+  SELECT 1 FROM information_schema.schemata WHERE schema_name = $1 LIMIT 1
+`;
+async function resolveInferenceLogConfig(pgClient, databaseId) {
+    try {
+        const schemaCheck = await pgClient.query(SCHEMA_EXISTS_SQL, ['metaschema_modules_public']);
+        if (schemaCheck.rows.length === 0)
+            return null;
+        const result = await pgClient.query(INFERENCE_LOG_MODULE_SQL, [databaseId]);
+        const row = result.rows[0];
+        if (!row?.schema || !row?.table_name)
+            return null;
+        return {
+            schema: row.schema,
+            tableName: row.table_name,
+        };
+    }
+    catch {
+        return null;
+    }
+}
+async function resolveBillingConfig(pgClient, databaseId) {
+    try {
+        // Guard: check if the metaschema_modules_public schema exists.
+        // If the database doesn't have the billing module provisioned,
+        // this schema (or the billing_module table) won't exist.
+        const schemaCheck = await pgClient.query(SCHEMA_EXISTS_SQL, ['metaschema_modules_public']);
+        if (schemaCheck.rows.length === 0)
+            return null;
+        const result = await pgClient.query(BILLING_MODULE_SQL, [databaseId]);
+        const row = result.rows[0];
+        if (!row?.record_usage_function)
+            return null;
+        return {
+            publicSchema: row.public_schema,
+            privateSchema: row.private_schema,
+            recordUsageFunction: row.record_usage_function,
+            // The check_billing_quota function name follows the inflection pattern
+            checkBillingQuotaFunction: 'check_billing_quota',
+        };
+    }
+    catch {
+        // Schema/table doesn't exist or query failed — billing not available
+        return null;
+    }
+}
+// ─── Public API ─────────────────────────────────────────────────────────────
+/**
+ * Resolve billing config for a database.
+ * Results are cached per database_id with a 5-minute TTL.
+ *
+ * @param pgClient - A client connected to the tenant database (from withPgClient)
+ * @param databaseId - The database UUID
+ */
+export async function getLlmBillingConfig(pgClient, databaseId) {
+    const cached = billingCache.get(databaseId);
+    if (cached)
+        return cached;
+    const [billing, inferenceLog] = await Promise.all([
+        resolveBillingConfig(pgClient, databaseId),
+        resolveInferenceLogConfig(pgClient, databaseId),
+    ]);
+    const entry = { billing, inferenceLog };
+    billingCache.set(databaseId, entry);
+    return entry;
+}
+/**
+ * Invalidate the cached config for a specific database (or all).
+ */
+export function invalidateLlmBillingConfig(databaseId) {
+    if (databaseId) {
+        billingCache.delete(databaseId);
+    }
+    else {
+        billingCache.clear();
+    }
+}
+/**
+ * Get cache stats for diagnostics.
+ */
+export function getLlmBillingCacheStats() {
+    return { size: billingCache.size, max: 50 };
+}

package/esm/embedder.d.ts

@@ -24,12 +24,12 @@ export declare function buildEmbedder(config: EmbedderConfig): EmbedderFunction
  */
 export declare function buildEmbedderFromModule(data: LlmModuleData): EmbedderFunction | null;
 /**
- * Resolve an embedder from environment variables via getEnvOptions().
+ * Resolve an embedder from environment variables.
  * This is a fallback for development when no llm_module or defaultEmbedder is configured.
  *
- * Environment variables (parsed by @constructive-io/graphql-env):
- *   EMBEDDER_PROVIDER - Provider name ('ollama')
- *   EMBEDDER_MODEL    - Model identifier
- *   EMBEDDER_BASE_URL - Provider base URL
+ * Environment variables (with defaults from env.ts):
+ *   EMBEDDER_PROVIDER  - Provider name (default: 'ollama')
+ *   EMBEDDER_MODEL     - Model identifier (default: 'nomic-embed-text')
+ *   EMBEDDER_BASE_URL  - Provider base URL (default: 'http://localhost:11434')
  */
 export declare function buildEmbedderFromEnv(): EmbedderFunction | null;

package/esm/embedder.js

@@ -10,7 +10,7 @@
  *   3. Environment variables (EMBEDDER_PROVIDER, EMBEDDER_MODEL, EMBEDDER_BASE_URL)
  */
 import OllamaClient from '@agentic-kit/ollama';
-import { getEnvOptions } from '@constructive-io/graphql-env';
+import { getLlmEnvOptions } from './env';
 // ─── Built-in Providers ─────────────────────────────────────────────────────
 /**
  * Create an Ollama-based embedder function.
@@ -48,26 +48,18 @@ export function buildEmbedderFromModule(data) {
         provider: data.embedding_provider,
         model: data.embedding_model,
         baseUrl: data.embedding_base_url,
-        apiKey: data.api_key_ref,
     });
 }
 /**
- * Resolve an embedder from environment variables via getEnvOptions().
+ * Resolve an embedder from environment variables.
  * This is a fallback for development when no llm_module or defaultEmbedder is configured.
  *
- * Environment variables (parsed by @constructive-io/graphql-env):
- *   EMBEDDER_PROVIDER - Provider name ('ollama')
- *   EMBEDDER_MODEL    - Model identifier
- *   EMBEDDER_BASE_URL - Provider base URL
+ * Environment variables (with defaults from env.ts):
+ *   EMBEDDER_PROVIDER  - Provider name (default: 'ollama')
+ *   EMBEDDER_MODEL     - Model identifier (default: 'nomic-embed-text')
+ *   EMBEDDER_BASE_URL  - Provider base URL (default: 'http://localhost:11434')
  */
 export function buildEmbedderFromEnv() {
-    const { llm } = getEnvOptions();
-    const provider = llm?.embedder?.provider;
-    if (!provider)
-        return null;
-    return buildEmbedder({
-        provider,
-        model: llm?.embedder?.model,
-        baseUrl: llm?.embedder?.baseUrl,
-    });
+    const { embedding } = getLlmEnvOptions();
+    return buildEmbedder(embedding);
 }

package/esm/env.d.ts

@@ -0,0 +1,31 @@
+/**
+ * LLM Environment Configuration
+ *
+ * Single source of truth for all LLM-related environment variables and defaults.
+ * Every other module in graphile-llm imports from here — no direct process.env
+ * reads or scattered null coalescing elsewhere.
+ *
+ * Environment variables:
+ *   EMBEDDER_PROVIDER  - Embedding provider name ('ollama')
+ *   EMBEDDER_MODEL     - Embedding model (default: 'nomic-embed-text')
+ *   EMBEDDER_BASE_URL  - Embedding provider URL (default: 'http://localhost:11434')
+ *   CHAT_PROVIDER      - Chat provider name ('ollama')
+ *   CHAT_MODEL         - Chat model (default: 'llama3')
+ *   CHAT_BASE_URL      - Chat provider URL (default: 'http://localhost:11434')
+ */
+export interface LlmProviderConfig {
+    provider: string;
+    model: string;
+    baseUrl: string;
+}
+export interface LlmEnvOptions {
+    embedding: LlmProviderConfig;
+    chat: LlmProviderConfig;
+}
+/**
+ * Resolve LLM configuration from environment variables with sensible defaults.
+ *
+ * Call this once and pass the result around — never read process.env directly
+ * in plugin code.
+ */
+export declare function getLlmEnvOptions(): LlmEnvOptions;

package/esm/env.js

@@ -0,0 +1,49 @@
+/**
+ * LLM Environment Configuration
+ *
+ * Single source of truth for all LLM-related environment variables and defaults.
+ * Every other module in graphile-llm imports from here — no direct process.env
+ * reads or scattered null coalescing elsewhere.
+ *
+ * Environment variables:
+ *   EMBEDDER_PROVIDER  - Embedding provider name ('ollama')
+ *   EMBEDDER_MODEL     - Embedding model (default: 'nomic-embed-text')
+ *   EMBEDDER_BASE_URL  - Embedding provider URL (default: 'http://localhost:11434')
+ *   CHAT_PROVIDER      - Chat provider name ('ollama')
+ *   CHAT_MODEL         - Chat model (default: 'llama3')
+ *   CHAT_BASE_URL      - Chat provider URL (default: 'http://localhost:11434')
+ */
+// ─── Defaults ───────────────────────────────────────────────────────────────
+const LLM_DEFAULTS = {
+    embedding: {
+        provider: 'ollama',
+        model: 'nomic-embed-text',
+        baseUrl: 'http://localhost:11434',
+    },
+    chat: {
+        provider: 'ollama',
+        model: 'llama3',
+        baseUrl: 'http://localhost:11434',
+    },
+};
+// ─── Resolution ─────────────────────────────────────────────────────────────
+/**
+ * Resolve LLM configuration from environment variables with sensible defaults.
+ *
+ * Call this once and pass the result around — never read process.env directly
+ * in plugin code.
+ */
+export function getLlmEnvOptions() {
+    return {
+        embedding: {
+            provider: process.env.EMBEDDER_PROVIDER ?? LLM_DEFAULTS.embedding.provider,
+            model: process.env.EMBEDDER_MODEL ?? LLM_DEFAULTS.embedding.model,
+            baseUrl: process.env.EMBEDDER_BASE_URL ?? LLM_DEFAULTS.embedding.baseUrl,
+        },
+        chat: {
+            provider: process.env.CHAT_PROVIDER ?? LLM_DEFAULTS.chat.provider,
+            model: process.env.CHAT_MODEL ?? LLM_DEFAULTS.chat.model,
+            baseUrl: process.env.CHAT_BASE_URL ?? LLM_DEFAULTS.chat.baseUrl,
+        },
+    };
+}

package/esm/index.d.ts

@@ -29,11 +29,20 @@
  * };
  * ```
  */
+export { getLlmEnvOptions } from './env';
+export type { LlmEnvOptions, LlmProviderConfig } from './env';
 export { GraphileLlmPreset } from './preset';
 export { createLlmModulePlugin } from './plugins/llm-module-plugin';
 export { createLlmTextSearchPlugin } from './plugins/text-search-plugin';
 export { createLlmTextMutationPlugin } from './plugins/text-mutation-plugin';
 export { createLlmRagPlugin } from './plugins/rag-plugin';
+export { createLlmMeteringPlugin } from './plugins/metering-plugin';
+export { getAgentDiscovery, clearAgentDiscoveryCache } from './plugins/agent-discovery-plugin';
+export type { AgentTableInfo, AgentDiscovery } from './plugins/agent-discovery-plugin';
 export { buildEmbedder, buildEmbedderFromModule, buildEmbedderFromEnv, } from './embedder';
 export { buildChatCompleter, buildChatCompleterFromModule, buildChatCompleterFromEnv, } from './chat';
-export type { EmbedderFunction, EmbedderConfig, ChatFunction, ChatConfig, ChatMessage, ChatOptions, LlmModuleData, GraphileLlmOptions, RagDefaults, ChunkTableInfo, } from './types';
+export { meteredEmbed, meteredChat, logInferenceUsage, QuotaExceededError } from './metering';
+export type { MeteringContext, MeteringOptions, MeterResult, WithPgClient, InferenceLogEntry } from './metering';
+export { getLlmBillingConfig, invalidateLlmBillingConfig, getLlmBillingCacheStats, } from './config-cache';
+export type { BillingConfig, LlmBillingCacheEntry, InferenceLogConfig, PgClient } from './config-cache';
+export type { EmbedderFunction, EmbedderConfig, ChatFunction, ChatConfig, ChatMessage, ChatOptions, LlmModuleData, GraphileLlmOptions, MeteringConfig, RagDefaults, ChunkTableInfo, } from './types';

package/esm/index.js

@@ -29,14 +29,24 @@
  * };
  * ```
  */
+// Environment configuration (single source of truth for LLM defaults)
+export { getLlmEnvOptions } from './env';
 // Preset (recommended entry point)
 export { GraphileLlmPreset } from './preset';
-// Individual plugins
+// Individual plugins (pure — no billing dependency)
 export { createLlmModulePlugin } from './plugins/llm-module-plugin';
 export { createLlmTextSearchPlugin } from './plugins/text-search-plugin';
 export { createLlmTextMutationPlugin } from './plugins/text-mutation-plugin';
 export { createLlmRagPlugin } from './plugins/rag-plugin';
+// Metering plugin (opt-in billing integration)
+export { createLlmMeteringPlugin } from './plugins/metering-plugin';
+// Agent discovery (queries agent_chat_module config table at runtime)
+export { getAgentDiscovery, clearAgentDiscoveryCache } from './plugins/agent-discovery-plugin';
 // Embedder utilities
 export { buildEmbedder, buildEmbedderFromModule, buildEmbedderFromEnv, } from './embedder';
 // Chat completion utilities
 export { buildChatCompleter, buildChatCompleterFromModule, buildChatCompleterFromEnv, } from './chat';
+// Metering utilities (for custom integration)
+export { meteredEmbed, meteredChat, logInferenceUsage, QuotaExceededError } from './metering';
+// Config cache (for custom integration)
+export { getLlmBillingConfig, invalidateLlmBillingConfig, getLlmBillingCacheStats, } from './config-cache';

package/esm/metering.d.ts

@@ -0,0 +1,114 @@
+/**
+ * metering — Billing-aware wrappers for embedder and chat functions
+ *
+ * Wraps EmbedderFunction and ChatFunction with:
+ *   1. Pre-check: `check_billing_quota(meter_slug, entity_id, estimated_amount)`
+ *   2. Execute the underlying function
+ *   3. Post-record: `record_usage(meter_slug, entity_id, actual_amount)`
+ *
+ * When the quota check fails, the wrapper returns null (graceful degradation)
+ * instead of throwing, so the search pipeline can fall back to text-only.
+ *
+ * Token counts are estimated from text length (~4 chars per token). No
+ * tokenizer needed — the billing system uses tokens as abstract units
+ * and the credit_cost on each model's meter normalizes the relative expense.
+ *
+ * The billing functions live in the tenant database and are called via the
+ * Graphile `withPgClient` callback. Function locations (schema, names) are
+ * resolved from `billing_module` metaschema and cached by `config-cache.ts`.
+ */
+import type { PgClient, BillingConfig, InferenceLogConfig } from './config-cache';
+import type { EmbedderFunction, ChatFunction, ChatMessage, ChatOptions } from './types';
+/**
+ * Callback matching Graphile's withPgClient signature.
+ * Acquires a pg client, calls the callback, then releases the client.
+ */
+export type WithPgClient = (pgSettings: Record<string, string>, callback: (pgClient: PgClient) => Promise<void>) => Promise<void>;
+export interface MeteringContext {
+    /** Callback to acquire a tenant database client */
+    withPgClient: WithPgClient;
+    /** pgSettings from the GraphQL context (for role/claims) */
+    pgSettings: Record<string, string>;
+    /** Billing function references from the billing_module */
+    billing: BillingConfig;
+    /** Entity ID to meter against (from JWT claims) */
+    entityId: string;
+    /** Per-request correlation ID (from request.id pgSetting) */
+    requestId: string | null;
+    /** Database UUID from JWT claims */
+    databaseId: string;
+    /** Actor (user) ID from JWT claims */
+    actorId: string | null;
+    /** Inference log table config (null if inference_log_module not provisioned) */
+    inferenceLog: InferenceLogConfig | null;
+}
+export interface MeteringOptions {
+    /** Meter slug for embedding operations (default: model name from build config) */
+    embeddingMeterSlug?: string;
+    /** Meter slug for chat completion operations (default: model name from build config) */
+    chatMeterSlug?: string;
+    /** Whether to skip metering entirely (e.g. for local dev). Default: false */
+    skipMetering?: boolean;
+    /** Embedding model name (for inference log) */
+    embeddingModel?: string;
+    /** Chat model name (for inference log) */
+    chatModel?: string;
+    /** Provider name (for inference log) */
+    provider?: string;
+}
+export interface MeterResult<T> {
+    /** The result from the underlying function, or null if quota exceeded */
+    result: T | null;
+    /** Whether the call was metered */
+    metered: boolean;
+    /** Whether the call was skipped due to quota limits */
+    quotaExceeded: boolean;
+    /** Latency of the underlying function call in ms */
+    latencyMs: number;
+}
+export interface InferenceLogEntry {
+    databaseId: string;
+    entityId: string;
+    actorId: string | null;
+    model: string;
+    provider: string | null;
+    service: 'llm' | 'embedding' | 'tts' | 'stt' | 'ocr' | 'image_gen' | 'search' | 'compute';
+    operation: string;
+    inputTokens: number;
+    outputTokens: number;
+    totalTokens: number;
+    cacheReadTokens: number | null;
+    cacheWriteTokens: number | null;
+    latencyMs: number;
+    ragEnabled: boolean;
+    chunksRetrieved: number | null;
+    embeddingModel: string | null;
+    embeddingLatencyMs: number | null;
+    status: 'success' | 'quota_exceeded' | 'provider_error' | 'timeout';
+    errorType: string | null;
+    rawUsage: Record<string, unknown> | null;
+}
+/**
+ * Write a row to the usage_log_inference table.
+ * Gracefully skips if the inference_log_module is not provisioned.
+ *
+ * TODO: Also write to child (generated) database when dual-write is needed.
+ */
+export declare function logInferenceUsage(ctx: MeteringContext, entry: InferenceLogEntry): Promise<void>;
+/**
+ * Wrap an embedder with billing quota check + usage recording.
+ *
+ * The returned MeterResult contains `quotaExceeded: true` when the pre-check
+ * fails, enabling the caller to fall back to text-only search.
+ */
+export declare function meteredEmbed(embedder: EmbedderFunction, text: string, ctx: MeteringContext | null, options?: MeteringOptions): Promise<MeterResult<number[]>>;
+/**
+ * Wrap a chat completion call with billing quota check + usage recording.
+ */
+export declare function meteredChat(chat: ChatFunction, messages: ChatMessage[], ctx: MeteringContext | null, chatOptions?: ChatOptions, meteringOptions?: MeteringOptions): Promise<MeterResult<string>>;
+export declare class QuotaExceededError extends Error {
+    readonly code = "QUOTA_EXCEEDED";
+    readonly meterSlug: string;
+    readonly entityId: string;
+    constructor(meterSlug: string, entityId: string);
+}