This document is a full export audit across the three ecosystems we support: AI SDK, LangChain (JS/TS), and LlamaIndex (TS). The goal is to map every major export surface to our adapter primitives, highlight mismatches, and list gaps.

This is a catalogue — not a commitment to implement every construct. Anything outside our adapter surface is explicitly called out.

The goal is to ensure Recipes are portable across these ecosystems.

AI SDK (ai + provider + provider-utils)

TIP

See the Models Adapter Page for implementation details.

Packages audited:

• ai (core helpers + UI utilities)
• @ai-sdk/provider (model interfaces + telemetry)
• @ai-sdk/provider-utils (messages/tools/schemas)
• @ai-sdk/langchain (UI message/stream adapter + LangSmith transport)
• @ai-sdk-tools/cache (tool caching helpers)
• @ai-sdk-tools/memory (working memory + history provider)
• @ai-sdk-tools/store (chat UI store + hooks)
• @ai-sdk/openai, @ai-sdk/anthropic, ollama-ai-provider-v2 (providers)

Key export categories:

• Model interfaces: LanguageModelV2/V3, EmbeddingModelV2/V3, ImageModelV2/V3, SpeechModelV2/V3, TranscriptionModelV2/V3, RerankingModelV3.
• Model helpers: generateText, generateObject, streamText, streamObject, embed, embedMany, generateImage, generateSpeech, transcribe.
• Messages + tools (provider-utils): ModelMessage, Tool, Schema, jsonSchema, zodSchema, content part helpers.
• UI/runtime wrappers (ai): chat UI, streaming UI, transport, UI message schemas.

Mapping to our primitives:

• Model (LanguageModelV1) → covered.
• Embedder (EmbeddingModelV1) → covered.
• ImageModel (ImageModelV3) → covered.
• SpeechModel (SpeechModelV3) → covered.
• TranscriptionModel (TranscriptionModelV3) → covered.
• Reranker (RerankingModelV3) → covered.
• Messages, Tools, Schema, PromptTemplate → covered.
• Memory → partial via @ai-sdk-tools/memory (working memory + history provider).
• Cache → partial via @ai-sdk-tools/cache CacheStore adapters (TTL best-effort).
• Store → out of scope (@ai-sdk-tools/store is a UI state layer).
• ChatTransport → covered via createAiSdkChatTransport (adapts our runtime to ChatTransport for useChat).
• UI Stream → covered via createAiSdkInteractionEventStream (direct UIMessage stream).
• Streaming → covered (normalized events across AI SDK, LangChain, LlamaIndex).

Gaps:

• Provider middleware hooks exist but no adapter-level tracing bridge.
• No checkpoint/event stream/interrupt abstractions in the AI SDK surface.

LangChain (JS/TS)

TIP

See the Retrieval and Tools adapter pages.

Packages audited:

• @langchain/core
• @langchain/community
• @langchain/openai
• @langchain/ollama
• @langchain/textsplitters
• @langchain/langgraph (checkpoints/interrupts)
• @langchain/langgraph-checkpoint

Core module inventory (from @langchain/core/*.d.ts):

• agents, runnables, tools, retrievers, vectorstores
• documents, document_loaders, memory, chat_history, stores
• prompts, prompt_values, messages, outputs, output_parsers
• structured_query, indexing, caches, callbacks, tracers

Mapping to our primitives:

• Model, Embedder, Retriever, Reranker, TextSplitter, Transformer, DocumentLoader, VectorStore, Memory, KVStore, Cache, Tool, PromptTemplate, OutputParser, StructuredQuery, Trace, EventStream, Indexing, Checkpoint → covered.
• Runnable, Agents, ChatHistory → not covered.

Gaps:

• Runnables/chains/agents are higher-level composition constructs.
• Callbacks/tracers map to trace sinks (custom + run.start/run.end + provider.response).

LlamaIndex (TS)

TIP

See the Retrieval Adapter Page for implementation details.

Packages audited (top-level modules):

• llms, embeddings, retriever, postprocessor, node-parser
• tools, prompts, memory, schema, vector-store, storage
• query-engine, response-synthesizers, chat-engine, agent, indices

Mapping to our primitives:

• Model, Embedder, Retriever, Reranker, TextSplitter, Transformer, DocumentLoader, VectorStore, Memory, KVStore, Cache, Tool, PromptTemplate → covered.
• QueryEngine, ResponseSynthesizer, EventStream, Checkpoint → covered.
• ChatEngine, Agent, Indices → not covered.

Gaps:

• Storage context and index lifecycle helpers are not modeled.
• Interrupt strategy remains core-only.

Media Construct Catalogue

AI SDK

• Image: ImageModelV3 (provider: @ai-sdk/provider)
• Speech (TTS): SpeechModelV3
• Transcription (STT): TranscriptionModelV3

Additional AI SDK interfaces worth tracking:

• RerankingModelV3 (provider) for rerankers.
• EmbeddingModelV3 and LanguageModelV3 (streaming-friendly).

LangChain (JS/TS)

• Image generation:
  • DallEAPIWrapper (JS) in @langchain/openai (tools/dalle)
  • imageGeneration tool spec (Responses API tool)
  • These are tool-style wrappers, not a generic ImageModel interface.
• Audio output:
  • ChatOpenAI supports audio output (model modalities + audio options)
  • This is a feature of the chat model, not a standalone SpeechModel.
• Transcription:
  • Document loaders for audio (e.g., OpenAIWhisperAudio, SonixAudioTranscriptionLoader)
  • These map to DocumentLoader rather than a formal TranscriptionModel.

LlamaIndex (TS)

• Image generation:
  • Tool specs exist (e.g., OpenAI image generation tool spec)
  • Exposed via tool calling, not a dedicated ImageModel.
• Multimodal:
  • Multimodal LLMs handle image inputs; still a Model surface, not a media model.
• Audio / transcription:
  • No canonical speech/transcription model interface surfaced in core packages.
• Trace:
  • Workflow trace plugins map to Trace adapters (handler start/end/error).

UI Interaction SDKs

TIP

See the Interaction guide for implementation details.

We provide Interaction Adapters that bridge our InteractionEvent stream (generated by the InteractionPrimitive) into the native data formats of popular UI libraries. This allows you to use our agent/workflow runtime with standard frontend components.

AI SDK UI (ai / @ai-sdk/ui-utils)

• Adapter: createAiSdkInteractionEventStream
• Function: Converts our events into AI SDK's UIMessage stream format (text parts, tool calls, tool results).
• Transport: createAiSdkChatTransport (mirrors LanguageModelV1 for use with streamText / useChat on the server).
• Helpers: toAiSdkUiMessageChunks, createAiSdkInteractionMapper, createAiSdkInteractionSink.

Assistant UI (@assistant-ui/react)

• Adapter: createAssistantUiInteractionEventStream
• Function: Converts our events into assistant-ui's structured command stream (text deltas, tool call deltas, rich args).
• Integration: Designed to work with useAssistant or custom runtimes.

NLUX (@nlux/core)

• Adapter: createNluxChatAdapter
• Function: Creates a fully compatible NLUX adapter object (batchText, streamText) from our interaction handle.
• Features: Supports streaming, tool calls, and distinct "user" vs "assistant" roles.

OpenAI ChatKit (@openai/chatkit)

• Adapter: createChatKitInteractionEventStream
• Function: Projects events into a CustomEvent stream (e.g. openai-chatkit:thread:message:delta) on a DOM element.
• Target: Intended for use with the <openai-chatkit> web component or similar custom elements.

Parity Matrix

Legend:

• full: adapter exists and aligns with the construct
• partial: adapter exists but semantics differ or are limited
• provider-specific: adapter only for certain providers/tools
• missing: no adapter or ecosystem surface

Construct	AI SDK	LangChain	LlamaIndex	Mix-and-match implications
Model	full	full	full	Safe to mix; Model is the canonical cross-ecosystem surface.
Embedder	full	full	full	Safe; embedding dimension may still vary per provider.
Retriever	missing	full	full	AI SDK has no retriever abstraction; must bring one from LC/LI.
Reranker	full	full	full	AI SDK supports reranking via RerankingModelV3.
TextSplitter	missing	full	full	AI SDK has no splitter; LC/LI supported.
Transformer	missing	full	full	AI SDK has no transformer; LC/LI supported.
Storage (KVStore)	missing	full	full	AI SDK has no storage adapter.
Cache	partial	partial	partial	AI SDK CacheStore + LC BaseStore + LI BaseKVStore; TTL best-effort.
Checkpoint	missing	full	full	LlamaIndex snapshots map via SnapshotData stores.
EventStream	missing	full	full	LangChain via callback handlers; LlamaIndex workflow events.
Interrupt	missing	full	missing	LangGraph interrupt semantics map to restart strategies.
Memory	partial	full	full	AI SDK has a memory provider (@ai-sdk-tools/memory), mapped.
Tools	full	full	full	Safe; tool schemas normalize across ecosystems.
PromptTemplate	full	full	full	AI SDK Prompt input type adapters.
OutputParser	missing	full	missing	LangChain parsers only; others use schema-based generation.
StructuredQuery	missing	full	missing	LangChain structured query IR; others lack a query model.
Indexing	missing	full	missing	LangChain index + record manager; no AI SDK/LI parity yet.
QueryEngine	missing	missing	full	LlamaIndex query engine; separate from "Model".
ResponseSynthesizer	missing	missing	full	LlamaIndex synthesizers; separate from "Model".
VectorStore	missing	full	partial	AI SDK has none; LI delete filters unsupported.
ImageModel	full	missing	missing	AI SDK V3 direct; LC/LI require tool wrappers.
SpeechModel	full	missing	missing	AI SDK V3 direct; LC via ChatOpenAI audio output.
TranscriptionModel	full	missing	missing	AI SDK V3 direct; LC via document loaders (shape mismatch).
Trace	missing	full	full	LangChain callbacks + LlamaIndex trace plugins map to trace sink.
ChatTransport	full	missing	missing	AI SDK ChatTransport adapter; others need custom implementations.
Streaming	full	full	full	Normalized stream events; provider-specific parts remain raw.

Summary: Coverage vs Gaps

Covered primitives (cross-ecosystem):

• Model, Embedder, Retriever, Reranker
• TextSplitter, Transformer, DocumentLoader
• VectorStore (write path), Memory (LC/LI), KVStore (LC/LI), Cache (partial)
• Tool, PromptTemplate, Schema, Messages
• Trace (LangChain + LlamaIndex)
• OutputParser (LangChain only)
• StructuredQuery (LangChain only)
• ImageModel, SpeechModel, TranscriptionModel (AI SDK direct; LC/LI provider-specific)

Gaps (candidate future primitives):

• Indexing (LangChain only).
• QueryEngine and ResponseSynthesizer (LlamaIndex only).