# LiteLLM local Ollama and OpenAI-compatible backends > 12 source-anchored research claims on LiteLLM local Ollama and OpenAI-compatible backends, grouped by facet. Original confidence and source-owner limits are retained. Parent: [LiteLLM gateway and SDK engineering](https://llms-explorer.com/tree/litellm-gateway-sdk-engineering/) · 5 facets · 12 facts · page: https://llms-explorer.com/tree/litellm-local-ollama-and-openai-compatible-backends/ ## Definitions - Native tool support is a model/backend capability. Ollama exposes tools and structured tool_calls; LiteLLM's documentation warns that not every Ollama model supports function calling. — [source](https://docs.ollama.com/api/chat) *(confidence medium; native contracts do not independently certify LiteLLM implementation · confidence: medium)* - Use ollama_chat/ for Ollama POST /api/chat; LiteLLM recommends that path rather than legacy ollama/ for conversational responses. — [source](https://docs.litellm.ai/docs/providers/ollama) *(confidence low; single-owner LiteLLM evidence; qualify exact deployment · confidence: low)* ## How it works - At the pinned commit, the Ollama Chat adapter forwards native tools directly and sets finish_reason=tool_calls when structured calls appear, including streamed calls. — [source](https://raw.githubusercontent.com/BerriAI/litellm/82d8b3797cf124e2baaa9c342f87a57fbb3a1a96/litellm/llms/ollama/chat/transformation.py) *(confidence low; single-owner LiteLLM evidence; qualify exact deployment · confidence: low)* ## Parameters and configuration - Native Ollama api_base points at the server root, such as http://localhost:11434. OpenAI-compatible calls need the backend's API root, commonly ending in /v1, not a full operation URL. — [source](https://docs.litellm.ai/docs/providers/openai_compatible) *(confidence medium; native contracts do not independently certify LiteLLM implementation · confidence: medium)* - vLLM automatic tool selection requires --enable-auto-tool-choice, a compatible --tool-call-parser and a tool-compatible chat template. The --chat-template argument is optional when tokenizer_config.json already supplies a suitable template. The cited Llama-3.1-8B-Instruct quickstart uses llama3_json and examples/tool_chat_template_llama3.1_json.jinja; those choices are specific to that example. Enabling an OpenAI-compatible HTTP endpoint alone does not configure these model-specific requirements. — [source](https://docs.vllm.ai/en/latest/features/tool_calling/) *(confidence low; single-owner vLLM evidence; qualify exact deployment · confidence: low)* - For Anthropic clients on a chat-only openai/ backend, opt into the Messages-to-Chat path. For clients sending Responses requests, use_chat_completions_api controls a separate Responses-to-Chat bridge. — [source](https://docs.litellm.ai/docs/proxy/config_settings) *(confidence low; single-owner LiteLLM evidence; qualify exact deployment · confidence: low)* - Ollama's OpenAI interface has no standard context-size control. Configure context through a Modelfile alias; native Ollama options such as num_ctx belong to a different API path. — [source](https://docs.ollama.com/api/openai-compatibility) *(confidence medium; native contracts do not independently certify LiteLLM implementation · confidence: medium)* ## How-to and procedures - For native Ollama streaming, accumulate thinking, content, and tool_calls and replay the complete assistant turn with tool results in the next request; processing text alone loses tool state. — [source](https://docs.ollama.com/capabilities/tool-calling) *(confidence low; single-owner Ollama evidence; qualify exact deployment · confidence: low)* - For a generic OpenAI-compatible Chat target, configure openai/, its api_base, and an upstream API key or required placeholder; gateway and upstream credentials are separate settings. — [source](https://docs.litellm.ai/docs/providers/openai_compatible) *(confidence medium; native contracts do not independently certify LiteLLM implementation · confidence: medium)* - Engineering validation should cover schemas, call/result pairing, parallel calls, streamed argument assembly, and follow-up turns. These are separate protocol obligations, not satisfied by a text-only smoke test. — [source](https://platform.claude.com/docs/en/agents-and-tools/tool-use/handle-tool-calls) *(confidence medium; native contracts do not independently certify LiteLLM implementation · confidence: medium)* ## Problems, failure modes and limitations - The pinned native Ollama Chat adapter removes tool_choice rather than enforcing it. Do not rely on forced named, required, or none behavior through this path without validation. — [source](https://raw.githubusercontent.com/BerriAI/litellm/82d8b3797cf124e2baaa9c342f87a57fbb3a1a96/litellm/llms/ollama/chat/transformation.py) *(confidence low; single-owner LiteLLM evidence; qualify exact deployment · confidence: low)* - Schema-constrained output can produce a parseable tool call without a good task decision. Validate semantic arguments and useful tool selection before treating a backend as research-capable. — [source](https://docs.vllm.ai/en/latest/features/tool_calling/) *(confidence low; single-owner vLLM evidence; qualify exact deployment · confidence: low)*