@marinoscar/kvox-eval 0.1.0-next.0 → 0.1.0-next.1

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # @marinoscar/kvox-eval
 
+## 0.1.0-next.1
+
+### Minor Changes
+
+- 38bf270: `@marinoscar/kvox-ai` and `@marinoscar/kvox-eval` no longer require `@marinoscar/platform-api`: they depend on the lean `@marinoscar/platform-ai-providers` (provider adapters and the provider-agnostic AI core, no Nest runtime) as an exact-version peer, and `kvox-eval` gains an `openai` peer for its live runs (#578). `kvox-ai` exports `isAiError()`, which recognises the lean `AiError` and the web host's distinct `AiError` class, so error classification is unchanged in both editions.
+
 ## 0.1.0-next.0
 
 ### Minor Changes

package/README.md

@@ -21,7 +21,7 @@ Entry points:
 ## Install and peer dependencies
 
 ```bash
-npm install @marinoscar/kvox-eval@next @marinoscar/kvox-ai@next @marinoscar/kvox-api@next @marinoscar/kvox-contract@next @marinoscar/platform-api@0.1.0-next.5 zod \
+npm install @marinoscar/kvox-eval@next @marinoscar/kvox-ai@next @marinoscar/kvox-api@next @marinoscar/kvox-contract@next @marinoscar/platform-ai-providers@0.1.0-next.6 openai zod \
   @electric-sql/pglite@0.5.10 @electric-sql/pglite-pgvector@0.0.11 @electric-sql/pglite-socket@0.2.13 \
   @prisma/client @prisma/adapter-pg pg @nestjs/common @nestjs/core reflect-metadata rxjs
 ```
@@ -34,7 +34,8 @@ Every slice is universal, so every peer is required, including those only `ask-r
 | `@marinoscar/kvox-contract` | `>=0.0.0-0 \|\| >=0.1.0-0` | `fixtures`, `kg` (the ontology registry), `ask` (the Ask wire types) |
 | `@marinoscar/kvox-api` | `>=0.0.0-0 \|\| >=0.1.0-0` | `kg` (the extraction pipeline, `graph-extraction` and `graph`), `live` and `cli` (`KvoxAiPort`, `KvoxAiError`) |
 | `@marinoscar/kvox-ai` | `>=0.0.0-0 \|\| >=0.1.0-0` | `live` (`createKvoxAi`, `createDirectRuntime`) |
-| `@marinoscar/platform-api` | `0.1.0-next.5` (exact, the repository's platform pin) | `live` and `cli`: the OpenAI and OpenAI-compatible adapters; `ask-run`: the provider event types of the replay. Only `@marinoscar/platform-api/ai`, as in `@marinoscar/kvox-ai`; an interim peer until the lean adapter package of #578. |
+| `@marinoscar/platform-ai-providers` | `0.1.0-next.6` (exact) | `live` and `cli`: `createOpenAiAdapter` and `createOpenAiCompatibleAdapter` from the `/openai` subpath; `parity`, `ask-run`: the provider event types. Only the package root and `/openai` are imported; `@marinoscar/platform-api` is not a peer and importing it is a violation. |
+| `openai` | `^7.23.0` | `live`: the OpenAI SDK that the lean `/openai` factories build their client from. Required, because `live` constructs the SDK client. |
 | `@electric-sql/pglite` | `0.5.10` (exact) | `ask-run`: the in-memory PostgreSQL |
 | `@electric-sql/pglite-pgvector` | `0.0.11` (exact) | `ask-run`: the `vector` extension the migrations create |
 | `@electric-sql/pglite-socket` | `0.2.13` (exact) | `ask-run`: the loopback socket the Prisma client connects to |
@@ -133,11 +134,11 @@ None: no log, metric or span (the Ask run's Nest context logs errors and warning
 
 ## Security notes
 
-- The API key is read from the environment only, never a flag. It is handed to `createDirectRuntime`'s `keys()` per call and never logged, printed, put in an error or written into a prediction or a report (a prediction carries `model` alone). A failed live call is reported by its `KvoxAiError` class, the platform error code and the HTTP status, never a message body.
+- The API key is read from the environment only, never a flag. It is handed to `createDirectRuntime`'s `keys()` per call and never logged, printed, put in an error or written into a prediction or a report (a prediction carries `model` alone). A failed live call is reported by its `KvoxAiError` class, the provider error code and the HTTP status, never a message body.
 - Real meetings never enter the repository: `--real-dir`, `--out` of a real-data run and `KG_EVAL_REAL_DIR` are refused (exit 3) when they resolve inside the git work tree, symlinks included (`assertOutsideRepo`).
 - The golden set is synthetic; its spec rejects a real-looking email address, a phone number and the product's name.
 - The Ask run uses an in-memory PGlite only: it has no connection string to a real database, so it cannot read or write one. The key never enters that database.
-- Apart from the `ask-run` slice (Prisma, Nest, PGlite), the source imports no Prisma, no `@app/shared`, and of the platform only `@marinoscar/platform-api/ai` (rule `host-neutral` of `npm run check:kvox-slice-peers`).
+- Apart from the `ask-run` slice (Prisma, Nest, PGlite), the source imports no Prisma, no `@app/shared`, and of the platform only `@marinoscar/platform-ai-providers` and its `/openai` subpath (rule `host-neutral` of `npm run check:kvox-slice-peers`).
 
 ## Conformance suite
 

package/dist/ask-run/replay.d.ts

@@ -1,5 +1,5 @@
 import type { AskMessageResponse } from '@marinoscar/kvox-contract/ask';
-import type { AiStreamEvent } from '@marinoscar/platform-api/ai';
+import type { AiStreamEvent } from '@marinoscar/platform-ai-providers';
 /**
  * The provider and model a replay names (the scripted runtime answers for them).
  *

package/dist/cli/kg-command.d.ts

@@ -1,4 +1,4 @@
-import type { AiProviderAdapter } from '@marinoscar/platform-api/ai';
+import type { AiProviderAdapter } from '@marinoscar/platform-ai-providers';
 /**
  * The exit codes of `kvox-eval kg` (and of the web app's `npm run kg:eval`).
  *

package/dist/live/adapters.d.ts

@@ -1,4 +1,4 @@
-import type { AiProviderAdapter } from '@marinoscar/platform-api/ai';
+import type { AiProviderAdapter } from '@marinoscar/platform-ai-providers';
 /**
  * A provider id {@link platformAdapter} can build.
  *
@@ -6,11 +6,11 @@ import type { AiProviderAdapter } from '@marinoscar/platform-api/ai';
  */
 export type PlatformAdapterProvider = 'openai' | 'openai-compatible';
 /**
- * The platform's adapter for a provider, constructed outside Nest.
+ * The lean provider package's adapter for a provider, constructed outside Nest.
  *
  * @param provider - `openai` or `openai-compatible`.
  * @returns the adapter (`responses`, `embeddings`, ...).
- * @throws Error when the installed platform no longer carries the adapter where this shim looks.
+ * @throws Error when the adapter built does not carry the id asked for.
  * @stability experimental
  */
 export declare function platformAdapter(provider: PlatformAdapterProvider): AiProviderAdapter;

package/dist/live/adapters.js

@@ -3,56 +3,31 @@
 // The provider adapters a live run calls (issue #580)
 // =============================================================================
 //
-// The OpenAI and OpenAI-compatible adapters are the platform's
-// (`@marinoscar/platform-api/ai`). Its public entry exports their Nest MODULES
-// (`OpenAiProviderModule`, `OpenAiCompatibleProviderModule`), not the adapter
-// classes, so the adapter is read off the module's own `exports` metadata and
-// constructed by hand over a fresh `AiProviderRegistry` and the module's client
-// factory (no Nest application is booted: the adapter registers itself only in
-// `onModuleInit`, which is never called here).
+// The OpenAI and OpenAI-compatible adapters are the lean provider package's
+// (`@marinoscar/platform-ai-providers/openai`): plain factories, no Nest, no
+// database. `platformAdapter` calls the factory for the provider with no options
+// (no logger, no tracer: a live run is a CLI); the key and base URL travel per call
+// in the adapter's call context, never here.
 //
-// PLATFORM-ADOPTION-SHIM(#578): until the lean provider-adapter package exports
-// the adapter classes (or factories), at which point this file becomes two
-// `new` expressions.
+// `openai` (the SDK) is an OPTIONAL peer of this package: it is loaded only when a
+// live run asks for one of these two providers, never by a scoring run.
 // =============================================================================
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.platformAdapter = platformAdapter;
-/** The platform module that carries each provider's adapter. */
-const ADAPTER_MODULES = {
-    openai: 'OpenAiProviderModule',
-    'openai-compatible': 'OpenAiCompatibleProviderModule',
-};
-/** `Reflect.getMetadata`, installed by `reflect-metadata` (the platform's Nest import loads it). */
-function metadata(key, target) {
-    const reflect = Reflect;
-    if (typeof reflect.getMetadata !== 'function')
-        throw new Error('reflect-metadata is not loaded');
-    return reflect.getMetadata(key, target);
-}
 /**
- * The platform's adapter for a provider, constructed outside Nest.
+ * The lean provider package's adapter for a provider, constructed outside Nest.
  *
  * @param provider - `openai` or `openai-compatible`.
  * @returns the adapter (`responses`, `embeddings`, ...).
- * @throws Error when the installed platform no longer carries the adapter where this shim looks.
+ * @throws Error when the adapter built does not carry the id asked for.
  * @stability experimental
  */
 function platformAdapter(provider) {
-    // Required lazily: the platform's AI entry loads Nest and the OpenAI SDK, which a
-    // scoring run (gold, a predictions directory) has no reason to load.
-    const ai = require('@marinoscar/platform-api/ai');
-    const moduleClass = ai[ADAPTER_MODULES[provider]];
-    if (moduleClass === undefined)
-        throw new Error(`@marinoscar/platform-api/ai has no ${ADAPTER_MODULES[provider]}`);
-    const exported = metadata('exports', moduleClass) ?? [];
-    const providers = metadata('providers', moduleClass) ?? [];
-    const Adapter = exported.find((c) => typeof c === 'function');
-    const Factory = providers.find((c) => typeof c === 'function' && c !== Adapter);
-    if (Adapter === undefined || Factory === undefined) {
-        throw new Error(`${ADAPTER_MODULES[provider]} does not carry an adapter and a client factory`);
-    }
-    const adapter = new Adapter(new ai.AiProviderRegistry(), new Factory());
+    // Required lazily: the subpath loads the OpenAI SDK, which a scoring run (gold, a
+    // predictions directory) has no reason to load.
+    const openai = require('@marinoscar/platform-ai-providers/openai');
+    const adapter = provider === 'openai' ? openai.createOpenAiAdapter() : openai.createOpenAiCompatibleAdapter();
     if (adapter.id !== provider)
-        throw new Error(`${ADAPTER_MODULES[provider]} built '${adapter.id}', expected '${provider}'`);
+        throw new Error(`The ${provider} adapter factory built '${adapter.id}', expected '${provider}'`);
     return adapter;
 }

package/dist/live/live-ai.d.ts

@@ -1,6 +1,6 @@
 import { type KvoxTaskModelSource } from '@marinoscar/kvox-ai';
 import type { KvoxAiPort } from '@marinoscar/kvox-api/ports';
-import type { AiProviderAdapter } from '@marinoscar/platform-api/ai';
+import type { AiProviderAdapter } from '@marinoscar/platform-ai-providers';
 import { type PlatformAdapterProvider } from './adapters.js';
 /**
  * The providers a live run may name. `copilot` is listed so it can be refused

package/host/prisma/kvox.lock

@@ -1,5 +1,5 @@
 {
-  "kvoxDbVersion": "0.1.0-next.0",
+  "kvoxDbVersion": "0.1.0-next.1",
   "lockVersion": 1,
   "migrations": [
     {

package/host/prisma/platform.lock

@@ -245,7 +245,7 @@
       "since": "1.0.0"
     }
   ],
-  "platformVersion": "0.1.0-next.5",
+  "platformVersion": "0.1.0-next.6",
   "rawSqlIndexes": [
     {
       "definition": "CREATE UNIQUE INDEX ask_messages_one_running_turn_uniq_idx ON public.ask_messages USING btree (conversation_id) WHERE ((role = 'assistant'::\"AskMessageRole\") AND (status = ANY (ARRAY['pending'::\"AskMessageStatus\", 'streaming'::\"AskMessageStatus\"])))",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@marinoscar/kvox-eval",
-  "version": "0.1.0-next.0",
+  "version": "0.1.0-next.1",
   "description": "The KVox knowledge-graph and Ask evaluation: the synthetic golden set, its schemas, the scorers, the report and the live runners, over the model port only (CommonJS, with the kvox-eval CLI).",
   "license": "MIT",
   "type": "commonjs",
@@ -73,11 +73,12 @@
     "@marinoscar/kvox-ai": ">=0.0.0-0 || >=0.1.0-0",
     "@marinoscar/kvox-api": ">=0.0.0-0 || >=0.1.0-0",
     "@marinoscar/kvox-contract": ">=0.0.0-0 || >=0.1.0-0",
-    "@marinoscar/platform-api": "0.1.0-next.5",
+    "@marinoscar/platform-ai-providers": "0.1.0-next.6",
     "@nestjs/common": "^11.1.12",
     "@nestjs/core": "^11.1.12",
     "@prisma/adapter-pg": "^7.8.0",
     "@prisma/client": "^7.8.0",
+    "openai": "^7.23.0",
     "pg": "^8.23.0",
     "reflect-metadata": "^0.2.2",
     "rxjs": "^7.8.1",
@@ -87,12 +88,12 @@
     "@electric-sql/pglite": "0.5.10",
     "@electric-sql/pglite-pgvector": "0.0.11",
     "@electric-sql/pglite-socket": "0.2.13",
-    "@marinoscar/kvox-ai": "0.1.0-next.0",
-    "@marinoscar/kvox-api": "0.1.0-next.0",
-    "@marinoscar/kvox-contract": "0.1.0-next.0",
-    "@marinoscar/kvox-db": "0.1.0-next.0",
-    "@marinoscar/platform-api": "0.1.0-next.5",
-    "@marinoscar/platform-db": "0.1.0-next.5",
+    "@marinoscar/kvox-ai": "0.1.0-next.1",
+    "@marinoscar/kvox-api": "0.1.0-next.1",
+    "@marinoscar/kvox-contract": "0.1.0-next.1",
+    "@marinoscar/kvox-db": "0.1.0-next.1",
+    "@marinoscar/platform-ai-providers": "0.1.0-next.6",
+    "@marinoscar/platform-db": "0.1.0-next.6",
     "@nestjs/common": "^11.1.12",
     "@nestjs/core": "^11.1.12",
     "@prisma/adapter-pg": "^7.8.0",
@@ -101,6 +102,7 @@
     "@types/jest": "^30.0.0",
     "@types/node": "^22.15.32",
     "jest": "^30.4.2",
+    "openai": "^7.23.0",
     "pg": "^8.23.0",
     "prisma": "^7.8.0",
     "reflect-metadata": "^0.2.2",