RLMCore · API

Defined in: packages/components/src/agents/rlm/RLMCore.ts:144

RLM Core - Recursive Language Model with REPL environment Based on the paper arXiv:2512.24601

Handles long contexts by allowing the LLM to generate code that orchestrates multiple sub-LLM calls for chunking and analysis tasks.

Supports true recursion: Sub-LLM can itself be an RLM instance with REPL capabilities.

Example

const rlm = new RLMCore(rootLLM, subLLM, 2); // depth=2 enables recursive RLM
const result = await rlm.executeStream({
  context: fileSystemContext,
  query: "Analyze this codebase architecture"
});

Implements

LLMProvider

Constructors

Constructor

new RLMCore(
   rootLLM, 
   subLLM, 
   maxDepth?, 
   sharedContext?, 
   customPrompt?, 
   options?
): RLMCore;

Defined in: packages/components/src/agents/rlm/RLMCore.ts:172

Create an RLM instance with optional recursive nesting

Parameters

Parameter	Type	Default value	Description
`rootLLM`	`LLMProvider`	`undefined`	LLM for generating orchestration code at this level
`subLLM`	`LLMProvider`	`undefined`	LLM for answering sub-queries (can be another RLM for recursion)
`maxDepth?`	`number`	`1`	Maximum nesting depth (default: 1)
`sharedContext?`	`FileSystemContext`	`undefined`	Shared FileSystemContext across all recursion levels (internal use)
`customPrompt?`	\| `string` \| `CustomPromptBuilder`	`undefined`	Additional task-specific prompt (appended to default)
`options?`	`RLMCoreOptions`	`{}`	Additional options, such as `verbose`

Returns

RLMCore

Remarks

When maxDepth > 1, the constructor automatically creates nested RLM instances:

depth=2: Root RLM → Sub RLM → Base LLM
depth=3: Root RLM → Sub RLM → Sub RLM → Base LLM

All nested RLMs share the same FileSystemContext (if provided) to avoid redundant file reads.

For most tasks, depth=1 is sufficient. Use a greater depth only for complex multi-layer analysis.

Methods

execute()

execute(input): Promise<RLMResult>;

Defined in: packages/components/src/agents/rlm/RLMCore.ts:199

Execute RLM with given context and query

Parameters

Parameter	Type
`input`	`RLMInput`

Returns

Promise<RLMResult>

executeStream()

executeStream(input): AsyncGenerator<RLMChunk>;

Defined in: packages/components/src/agents/rlm/RLMCore.ts:220

Execute RLM with streaming output

Parameters

Parameter	Type
`input`	`RLMInput`

Returns

AsyncGenerator<RLMChunk>

generate()

generate(prompt): Promise<string>;

Defined in: packages/components/src/agents/rlm/RLMCore.ts:708

Implement LLMProvider.generate - allows RLM to be used as sub-LLM This enables true recursive RLM where Sub-LLM is itself an RLM instance

Parameters

Parameter	Type
`prompt`	`string`

Returns

Promise<string>

Implementation of

LLMProvider.generate

generateStream()

generateStream(prompt): AsyncGenerator<string>;

Defined in: packages/components/src/agents/rlm/RLMCore.ts:718

Implement LLMProvider.generateStream - streaming variant for recursive RLM

Parameters

Parameter	Type
`prompt`	`string`

Returns

AsyncGenerator<string>

Implementation of

LLMProvider.generateStream