Python LiteLLM Provider Parity

This is the parity ledger for the providers currently registered by clojure-llm-sdk. It is intentionally not a Python LiteLLM provider-count target; unrelated providers are not added solely because the Python project supports them.

Baselines:

• Python BerriAI/LiteLLM supplies the provider catalog and endpoint-family vocabulary used by the ledger below.
• The separate Clojure library unravel-team/litellm-clj at 14bcdd949c0207d6c4988a3db887a1a7fa1c5522 is used only for the scoped implementation comparison in the next section.
• SDK scope is limited to providers exposed through llm.sdk/list-providers plus caller-registered Azure deployment profiles.

Scoped Clojure Library Comparison

At the pinned litellm-clj commit, its direct provider code already exposed Z.AI chat, Gemini embeddings, and Azure deployment embeddings. This SDK added the corresponding missing provider and surfaces: :zai, embeddings on the existing :gemini-native profile, and embeddings on Azure deployment profiles. Azure profiles also accept the current explicit :api-style :v1 route while retaining the classic API-version route as the default.

Three implementation principles were borrowed and adapted rather than copying that library's public API:

• OpenAI-compatible provider-native extras are flattened into the final wire body while transport-owned routing and canonical fields remain protected;
• assistant reasoning and tool state needed for a later turn remain replayable; and
• provider failures become structured, classified errors with retry semantics.

This SDK retains its richer namespaced canonical requests and typed response parts, including provider replay data, citations, reasoning, safety details, usage, cache, and cost. Provider-reported usage and parser-reported cost remain authoritative. Missing usage or pricing remains explicitly unknown and is never invented as zero. litellm-clj is not an SDK dependency, and no generated provider client was added.

SDK Surface Mapping

SDK surface	Python LiteLLM endpoint family
complete, streaming, tools, json-schema, reasoning	chat_completions
embedding	embeddings
rerank	rerank
image-generation	image_generations
transcription	audio_transcriptions
tts	audio_speech
moderation	moderations

Python LiteLLM endpoint families with no public SDK surface today are not implemented parity: raw provider-native messages/responses pass-through, a2a, interactions, batches, file lifecycle APIs, count_tokens, search, ocr, realtime, vector_stores, assistants, fine_tuning, rag, containers, image edits, and image variations. If those become SDK goals, they need first-class public functions, transports, schemas, fixtures, and live probes.

File/document attachment is different from file lifecycle management. The canonical :part/type :file message part is an SDK surface and is serialized for transports with provider-native file/document input support:

• OpenAI Responses / Codex: input_file
• Anthropic Messages: document content block
• Gemini / Vertex Gemini: fileData or inlineData
• Bedrock Converse: document content block

OpenAI Chat Completions and Cohere chat reject :file parts explicitly rather than stringifying maps into prompts.

Image-generation parity means a constructor exists, not that the SDK selects a billable model. OpenAI, OpenRouter, and Bedrock callers must provide an explicit :image/model; model availability and account access remain provider concerns.

Current Provider Ledger

SDK provider	Python LiteLLM provider	Constructor-backed SDK parity	Remaining Python LiteLLM gap for this provider family
:openai	openai	chat, streaming, tools, JSON schema, reasoning, file attachments on Chat Completions, embeddings, moderation, image generation, transcription, TTS	general Responses public surface, text completions, image edits/variations, realtime, assistants, vector stores, batches, fine tuning, containers, RAG
caller-defined Azure deployment	azure	chat, streaming, tools, JSON schema, reasoning, embeddings; classic deployment routing by API version or explicit /openai/v1 routing	No further Azure surface is claimed without a dedicated constructor and fixture evidence
:anthropic	anthropic	messages-backed chat, streaming, tools, JSON schema, thinking, OAuth/OAT, file/document attachments, native cache markers	count tokens, batches, file lifecycle APIs, skills, public raw messages/responses pass-through
:bedrock	bedrock	Converse chat, streaming eventstream, tools, JSON schema, reasoning, guardrails, cachePoint, multimodal file/document attachments, image generation, rerank	embeddings, count tokens, vector search/RAG, and provider-specific invoke transforms outside the implemented image families
:cohere	cohere	native v2 chat, streaming, tools, JSON schema, reasoning, citations, textual file/document attachments via documents, v2 embeddings, v2 rerank	responses/messages wrappers and a2a/interactions are not SDK surfaces
:gemini-native	gemini	native chat, streaming, tools, multimodal, file/document attachments, reasoning, cachedContent handling, native batch embeddings	count tokens, file lifecycle APIs, vector search/RAG, realtime, public generateContent pass-through
:vertex-gemini	vertex_ai	Vertex Gemini chat, streaming, tools, multimodal, file/document attachments, reasoning	embeddings, TTS, OCR, count tokens, fine tuning, RAG/vector stores, realtime, public generateContent pass-through
:vertex-anthropic	vertex_ai	Claude-on-Vertex messages chat, streaming, tools, JSON schema, thinking, file/document attachments, native cache markers (reuses Anthropic body/parse with GCP OAuth)	count tokens, batches, file lifecycle APIs, model listing, public rawPredict pass-through
:vertex-imagen	vertex_ai	Gemini image generation over Vertex generateContent (compatibility provider id)	image edits/videos and other Vertex AI endpoint families
:openrouter	openrouter	chat, streaming, tools, JSON schema, reasoning, provider routing, embeddings, image generation, live model/pricing lookup	responses, image edit, messages/a2a/interactions
:perplexity	perplexity	native Agent chat, streaming, tools, typed search output/citations, usage, and structured errors	dedicated search endpoint and other endpoint families not exposed through the Agent constructor
:codex	openai	OpenAI Responses-shaped Codex chat, streaming, tools, file attachments, reasoning, encrypted reasoning	not a general OpenAI Responses API surface
:codex-backend	chatgpt	ChatGPT backend Responses/SSE path, OAuth auth-json cache, tools, file attachments, reasoning	not a general ChatGPT automation surface
:ollama-native	ollama	native chat, streaming, tools, JSON schema, reasoning, multimodal input, embeddings	responses/messages/a2a/interactions wrappers
:voyage	voyage	dedicated text embeddings and rerank transports	contextualized and multimodal embedding endpoints require distinct canonical input shapes
:jina	jina_ai	dedicated dense/base64 embeddings and rerank transports	sparse, multivector, and multimodal outputs require distinct canonical result shapes
:elevenlabs	elevenlabs	TTS	Python LiteLLM also declares chat/messages/responses; this SDK treats ElevenLabs as a TTS provider only
:mistral	mistral	OpenAI-compatible chat, streaming, tools, JSON schema, reasoning, embeddings	OCR and audio transcription
:groq	groq	OpenAI-compatible chat, streaming, tools, JSON schema, reasoning, transcription	responses/messages/a2a/interactions
:deepseek	deepseek	OpenAI-compatible chat, streaming, tools, reasoning	responses/messages/a2a/interactions
:kimi	moonshot	OpenAI-compatible chat, streaming, tools, reasoning	responses/messages/a2a/interactions
:kimi-code	anthropic	Kimi Code OpenAI-compatible coding endpoint, streaming, tools, reasoning	Anthropic endpoint families are not implied by Kimi Code
:cerebras	cerebras	OpenAI-compatible chat, streaming, tools, JSON schema, reasoning	responses/messages/a2a/interactions
:together	together_ai	OpenAI-compatible chat, streaming, tools, JSON schema, reasoning, embeddings	responses/messages/a2a/interactions
:xai	xai	OpenAI-compatible chat, streaming, tools, JSON schema, reasoning	realtime and responses/messages/a2a/interactions
:zai	zai	OpenAI-compatible GLM chat, streaming, tools, reasoning and reasoning replay, multimodal image input, cached-token usage, structured errors	No additional Z.AI endpoint family is claimed without a dedicated constructor and fixture evidence
:huggingface	huggingface	OpenAI-compatible chat, streaming, tools, JSON schema	embeddings and rerank
:sambanova	sambanova	OpenAI-compatible chat, streaming, tools, JSON schema, reasoning	responses/messages/a2a/interactions
:deepinfra	deepinfra	OpenAI-compatible chat, streaming, tools, JSON schema, reasoning	responses/messages/a2a/interactions
:nebius	nebius	OpenAI-compatible chat, streaming, tools, JSON schema, embeddings	responses/messages/a2a/interactions
:hyperbolic	hyperbolic	OpenAI-compatible chat, streaming, tools	responses/messages/a2a/interactions
:novita	novita	OpenAI-compatible chat, streaming, tools, JSON schema	responses/messages/a2a/interactions
:friendliai	friendliai	OpenAI-compatible chat, streaming, tools	responses/messages/a2a/interactions
:featherless	featherless_ai	OpenAI-compatible chat, streaming, tools	responses/messages/a2a/interactions
:cloudflare	cloudflare	OpenAI-compatible chat, streaming, tools with caller-supplied account base URL	responses/messages/a2a/interactions; account-scoped model listing
:dashscope	dashscope	OpenAI-compatible chat, streaming, tools, JSON schema	responses/messages/a2a/interactions
:volcengine	volcengine	OpenAI-compatible chat, streaming, tools, JSON schema	responses/messages/a2a/interactions
:fake	none	deterministic test transport only	excluded from Python LiteLLM parity

Hard Rule

Provider coverage must stay constructor-backed. A provider row may not claim an SDK surface unless the profile has the matching transport constructor and the request can be built offline by llm.sdk.provider-coverage-test.