NVIDIA-NeMo · nabinchha · Mar 24, 2026 · Mar 24, 2026 · Mar 24, 2026 · Mar 24, 2026
@@ -7,7 +7,7 @@ This file provides guidance to agents when working with code in this repository.
 **DataDesigner** is an NVIDIA NeMo project for creating synthetic datasets from scratch. It's a comprehensive framework that generates structured data using multiple generation strategies:
 
 - **Sampled data**: Built-in generators (UUID, DateTime, etc.) and Faker integration
-- **LLM-generated content**: Text, code, and structured data via LiteLLM
+- **LLM-generated content**: Text, code, and structured data via native HTTP adapters
 - **Expression-based columns**: Derived columns using Jinja2 templates
 - **Validation & scoring**: Python, SQL, and remote validators; LLM-based judge scoring
 - **Seed dataset-based generation**: Generate from existing datasets
@@ -25,7 +25,7 @@ The project follows a layered architecture:
 2. **Engine Layer** ([packages/data-designer-engine/src/data_designer/engine/](packages/data-designer-engine/src/data_designer/engine/)): Internal generation and processing
    - `column_generators/`: Generates individual columns from configs
    - `dataset_builders/`: Orchestrates full dataset generation with DAG-based dependency management
-   - `models/`: LLM integration via LiteLLM with response parsing
+   - `models/`: LLM integration via native HTTP adapters with response parsing
    - `validators/`: Column validation (Python, SQL, Code, Remote)
    - `sampling_gen/`: Sophisticated person/entity sampling
 

@@ -14,11 +14,24 @@ The `ModelProvider` class has the following fields:
 |-------|------|----------|-------------|
 | `name` | `str` | Yes | Unique identifier for the provider (e.g., `"nvidia"`, `"openai"`, `"openrouter"`) |
 | `endpoint` | `str` | Yes | API endpoint URL (e.g., `"https://integrate.api.nvidia.com/v1"`) |
-| `provider_type` | `str` | No | Provider type (default: `"openai"`). Uses OpenAI-compatible API format |
+| `provider_type` | `str` | No | Provider type: `"openai"` (default) or `"anthropic"`. See [Supported Provider Types](#supported-provider-types) below |
 | `api_key` | `str` | No | API key or environment variable name (e.g., `"NVIDIA_API_KEY"`) |
 | `extra_body` | `dict[str, Any]` | No | Additional parameters to include in the request body of all API requests to the provider. |
 | `extra_headers` | `dict[str, str]` | No | Additional headers to include in all API requests to the provider. |
 
+## Supported Provider Types
+
+Data Designer supports two provider types:
+
+| Type | Description |
+|------|-------------|
+| `"openai"` | OpenAI-compatible chat completion API. This is the default and works with most providers, including NVIDIA NIM, vLLM, TGI, OpenRouter, Together AI, and OpenAI itself. |
+| `"anthropic"` | Anthropic's native Messages API for Claude models. Use this when connecting directly to Anthropic's API. |
+
+Most self-hosted and third-party endpoints expose an OpenAI-compatible API, so `provider_type="openai"` is the right choice in the majority of cases. Only use `"anthropic"` when connecting directly to Anthropic's API at `https://api.anthropic.com`.
+
+> **Note:** Previous versions of Data Designer supported additional provider types (e.g., `"azure"`, `"bedrock"`, `"vertex_ai"`) via a LiteLLM bridge. These are no longer supported. If you were using one of these types, switch to `provider_type="openai"` and point the `endpoint` to an OpenAI-compatible proxy or gateway for that service.
+
 ## API Key Configuration
 
 The `api_key` field can be specified in two ways:

@@ -115,7 +115,7 @@ def inject_sampler_type_into_params(cls, data: dict) -> dict:
 class LLMTextColumnConfig(SingleColumnConfig):
     """Configuration for text generation columns using Large Language Models.
 
-    LLM text columns generate free-form text content using language models via LiteLLM.
+    LLM text columns generate free-form text content using language models.
     Prompts support Jinja2 templating to reference values from other columns, enabling
     context-aware generation. The generated text can optionally include message traces
     capturing the full conversation history.

@@ -32,7 +32,6 @@
     "pq": "pyarrow.parquet",
     "pa": "pyarrow",
     "faker": "faker",
-    "litellm": "litellm",
     "sqlfluff": "sqlfluff",
     "httpx": "httpx",
     "duckdb": "duckdb",

@@ -2,7 +2,7 @@
 
 Generation engine for NeMo Data Designer synthetic data generation framework.
 
-This package contains the execution engine that powers Data Designer. It depends on `data-designer-config` and includes heavy dependencies like pandas, numpy, and LLM integration via litellm.
+This package contains the execution engine that powers Data Designer. It depends on `data-designer-config` and includes heavy dependencies like pandas, numpy, and native HTTP-based LLM integration.
 
 ## Installation
 

@@ -44,7 +44,6 @@ dependencies = [
     "json-repair>=0.48.0,<1",
     "jsonpath-rust-bindings>=1.0,<2",
     "jsonschema>=4.0.0,<5",
-    "litellm>=1.77.0,<1.80.12",
     "lxml>=6.0.2,<7",
     "marko>=2.1.2,<3",
     "mcp>=1.26.0,<2",

@@ -20,11 +20,12 @@
                                                               (TaskGroup)
 
 Singleton Event Loop:
-    The background loop is a process-wide singleton. LiteLLM and similar
-    libraries bind internal async state to a specific event loop, so creating
-    per-call or per-instance loops breaks connection reuse and triggers
-    cross-loop errors. ``ensure_async_engine_loop()`` creates one daemon
-    loop thread and reuses it for all executor instances.
+    The background loop is a process-wide singleton. Async-stateful
+    resources (connection pools, semaphores) bind internal state to a
+    specific event loop, so creating per-call or per-instance loops breaks
+    connection reuse and triggers cross-loop errors.
+    ``ensure_async_engine_loop()`` creates one daemon loop thread and
+    reuses it for all executor instances.
 
 Startup Handshake:
     Loop creation uses a ``threading.Event`` readiness handshake. The
@@ -90,8 +91,8 @@ def ensure_async_engine_loop() -> asyncio.AbstractEventLoop:
     """Get or create a persistent event loop for async engine work.
 
     A single event loop is shared across all AsyncConcurrentExecutor instances
-    to avoid breaking libraries (like LiteLLM) that bind internal async state
-    to a specific event loop.
+    to avoid breaking async-stateful resources (connection pools, semaphores)
+    that bind internal state to a specific event loop.
     """
     global _loop, _thread
     with _lock:

@@ -4,7 +4,6 @@
 from __future__ import annotations
 
 from data_designer.engine.models.clients.adapters.anthropic import AnthropicClient
-from data_designer.engine.models.clients.adapters.litellm_bridge import LiteLLMBridgeClient, LiteLLMRouter
 from data_designer.engine.models.clients.adapters.openai_compatible import OpenAICompatibleClient
 
-__all__ = ["AnthropicClient", "LiteLLMBridgeClient", "LiteLLMRouter", "OpenAICompatibleClient"]
+__all__ = ["AnthropicClient", "OpenAICompatibleClient"]
@@ -130,9 +130,9 @@ def map_http_error_to_provider_error(
 
 
 def extract_message_from_exception_string(raw: str) -> str:
-    """Extract a human-readable message from a stringified LiteLLM exception.
+    """Extract a human-readable message from a stringified provider exception.
 
-    LiteLLM often formats errors as ``"Error code: 400 - {json}"``.  This
+    Some providers format errors as ``"Error code: 400 - {json}"``.  This
     mirrors the structured-key lookup in ``_extract_structured_message`` but
     operates on a raw string instead of an ``HttpResponse``.
     """