Skip to content

Llm compiler

LLMCompilerAgentWorker #

Bases: BaseAgentWorker

LLMCompiler Agent Worker.

LLMCompiler is an agent framework that allows async multi-function calling and query planning. Here is the implementation.

Source Repo (paper linked): https://github.com/SqueezeAILab/LLMCompiler?tab=readme-ov-file

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
class LLMCompilerAgentWorker(BaseAgentWorker):
    """LLMCompiler Agent Worker.

    LLMCompiler is an agent framework that allows async multi-function calling and query planning.
    Here is the implementation.

    Source Repo (paper linked): https://github.com/SqueezeAILab/LLMCompiler?tab=readme-ov-file

    """

    def __init__(
        self,
        tools: Sequence[BaseTool],
        llm: LLM,
        callback_manager: Optional[CallbackManager] = None,
        verbose: bool = False,
        tool_retriever: Optional[ObjectRetriever[BaseTool]] = None,
        planner_example_prompt_str: Optional[str] = None,
        stop: Optional[List[str]] = None,
        joiner_prompt: Optional[PromptTemplate] = None,
        max_replans: int = 3,
    ) -> None:
        self.callback_manager = callback_manager or llm.callback_manager

        self.planner_example_prompt_str = (
            planner_example_prompt_str or PLANNER_EXAMPLE_PROMPT
        )
        self.system_prompt = generate_llm_compiler_prompt(
            tools, example_prompt=self.planner_example_prompt_str
        )
        self.system_prompt_replan = generate_llm_compiler_prompt(
            tools, is_replan=True, example_prompt=self.planner_example_prompt_str
        )

        self.llm = llm
        # TODO: make tool_retriever work
        self.tools = tools
        self.output_parser = LLMCompilerPlanParser(tools=tools)
        self.stop = stop
        self.max_replans = max_replans
        self.verbose = verbose

        # joiner program
        self.joiner_prompt = joiner_prompt or PromptTemplate(OUTPUT_PROMPT)
        self.joiner_program = LLMTextCompletionProgram.from_defaults(
            output_parser=LLMCompilerJoinerParser(),
            output_cls=JoinerOutput,
            prompt=self.joiner_prompt,
            llm=self.llm,
            verbose=verbose,
        )

        # if len(tools) > 0 and tool_retriever is not None:
        #     raise ValueError("Cannot specify both tools and tool_retriever")
        # elif len(tools) > 0:
        #     self._get_tools = lambda _: tools
        # elif tool_retriever is not None:
        #     tool_retriever_c = cast(ObjectRetriever[BaseTool], tool_retriever)
        #     self._get_tools = lambda message: tool_retriever_c.retrieve(message)
        # else:
        #     self._get_tools = lambda _: []

    @classmethod
    def from_tools(
        cls,
        tools: Optional[Sequence[BaseTool]] = None,
        tool_retriever: Optional[ObjectRetriever[BaseTool]] = None,
        llm: Optional[LLM] = None,
        callback_manager: Optional[CallbackManager] = None,
        verbose: bool = False,
        **kwargs: Any,
    ) -> "LLMCompilerAgentWorker":
        """Convenience constructor method from set of of BaseTools (Optional).

        Returns:
            LLMCompilerAgentWorker: the LLMCompilerAgentWorker instance

        """
        llm = llm or OpenAI(model=DEFAULT_MODEL_NAME)
        if callback_manager is not None:
            llm.callback_manager = callback_manager
        return cls(
            tools=tools or [],
            tool_retriever=tool_retriever,
            llm=llm,
            callback_manager=callback_manager,
            verbose=verbose,
        )

    def initialize_step(self, task: Task, **kwargs: Any) -> TaskStep:
        """Initialize step from task."""
        sources: List[ToolOutput] = []
        # temporary memory for new messages
        new_memory = ChatMemoryBuffer.from_defaults()

        # put user message in memory
        new_memory.put(ChatMessage(content=task.input, role=MessageRole.USER))

        # initialize task state
        task_state = {
            "sources": sources,
            "new_memory": new_memory,
        }
        task.extra_state.update(task_state)

        return TaskStep(
            task_id=task.task_id,
            step_id=str(uuid.uuid4()),
            input=task.input,
            step_state={"is_replan": False, "contexts": [], "replans": 0},
        )

    def get_tools(self, input: str) -> List[AsyncBaseTool]:
        """Get tools."""
        # return [adapt_to_async_tool(t) for t in self._get_tools(input)]
        return [adapt_to_async_tool(t) for t in self.tools]

    async def arun_llm(
        self,
        input: str,
        previous_context: Optional[str] = None,
        is_replan: bool = False,
    ) -> ChatResponse:
        """Run LLM."""
        if is_replan:
            system_prompt = self.system_prompt_replan
            assert previous_context is not None, "previous_context cannot be None"
            human_prompt = f"Question: {input}\n{previous_context}\n"
        else:
            system_prompt = self.system_prompt
            human_prompt = f"Question: {input}"

        messages = [
            ChatMessage(role=MessageRole.SYSTEM, content=system_prompt),
            ChatMessage(role=MessageRole.USER, content=human_prompt),
        ]

        return await self.llm.achat(messages)

    async def ajoin(
        self,
        input: str,
        tasks: Dict[int, LLMCompilerTask],
        is_final: bool = False,
    ) -> JoinerOutput:
        """Join answer using LLM/agent."""
        agent_scratchpad = "\n\n"
        agent_scratchpad += "".join(
            [
                task.get_thought_action_observation(
                    include_action=True, include_thought=True
                )
                for task in tasks.values()
                if not task.is_join
            ]
        )
        agent_scratchpad = agent_scratchpad.strip()

        output = self.joiner_program(
            query_str=input,
            context_str=agent_scratchpad,
        )
        output = cast(JoinerOutput, output)
        if self.verbose:
            print_text(f"> Thought: {output.thought}\n", color="pink")
            print_text(f"> Answer: {output.answer}\n", color="pink")
        if is_final:
            output.is_replan = False
        return output

    def _get_task_step_response(
        self,
        task: Task,
        llmc_tasks: Dict[int, LLMCompilerTask],
        answer: str,
        joiner_thought: str,
        step: TaskStep,
        is_replan: bool,
    ) -> TaskStepOutput:
        """Get task step response."""
        agent_answer = AgentChatResponse(response=answer, sources=[])

        if not is_replan:
            # generate final answer
            new_steps = []

            # put in memory
            task.extra_state["new_memory"].put(
                ChatMessage(content=answer, role=MessageRole.ASSISTANT)
            )
        else:
            # Collect contexts for the subsequent replanner
            context = generate_context_for_replanner(
                tasks=llmc_tasks, joiner_thought=joiner_thought
            )
            new_contexts = step.step_state["contexts"] + [context]
            # TODO: generate new steps
            new_steps = [
                step.get_next_step(
                    step_id=str(uuid.uuid4()),
                    input=None,
                    step_state={
                        "is_replan": is_replan,
                        "contexts": new_contexts,
                        "replans": step.step_state["replans"] + 1,
                    },
                )
            ]

        return TaskStepOutput(
            output=agent_answer,
            task_step=step,
            next_steps=new_steps,
            is_last=not is_replan,
        )

    async def _arun_step(
        self,
        step: TaskStep,
        task: Task,
    ) -> TaskStepOutput:
        """Run step."""
        if self.verbose:
            print(
                f"> Running step {step.step_id} for task {task.task_id}.\n"
                f"> Step count: {step.step_state['replans']}"
            )
        is_final_iter = (
            step.step_state["is_replan"]
            and step.step_state["replans"] >= self.max_replans
        )

        if len(step.step_state["contexts"]) == 0:
            formatted_contexts = None
        else:
            formatted_contexts = format_contexts(step.step_state["contexts"])
        llm_response = await self.arun_llm(
            task.input,
            previous_context=formatted_contexts,
            is_replan=step.step_state["is_replan"],
        )
        if self.verbose:
            print_text(f"> Plan: {llm_response.message.content}\n", color="pink")

        # return task dict (will generate plan, parse into dictionary)
        task_dict = self.output_parser.parse(cast(str, llm_response.message.content))

        # execute via task executor
        task_fetching_unit = TaskFetchingUnit.from_tasks(
            task_dict, verbose=self.verbose
        )
        await task_fetching_unit.schedule()

        ## join tasks - get response
        tasks = cast(Dict[int, LLMCompilerTask], task_fetching_unit.tasks)
        joiner_output = await self.ajoin(
            task.input,
            tasks,
            is_final=is_final_iter,
        )

        # get task step response (with new steps planned)
        return self._get_task_step_response(
            task,
            llmc_tasks=tasks,
            answer=joiner_output.answer,
            joiner_thought=joiner_output.thought,
            step=step,
            is_replan=joiner_output.is_replan,
        )

    @trace_method("run_step")
    def run_step(self, step: TaskStep, task: Task, **kwargs: Any) -> TaskStepOutput:
        """Run step."""
        return asyncio.run(self.arun_step(step=step, task=task, **kwargs))

    @trace_method("run_step")
    async def arun_step(
        self, step: TaskStep, task: Task, **kwargs: Any
    ) -> TaskStepOutput:
        """Run step (async)."""
        return await self._arun_step(step, task)

    @trace_method("run_step")
    def stream_step(self, step: TaskStep, task: Task, **kwargs: Any) -> TaskStepOutput:
        """Run step (stream)."""
        # # TODO: figure out if we need a different type for TaskStepOutput
        # return self._run_step_stream(step, task)
        raise NotImplementedError

    @trace_method("run_step")
    async def astream_step(
        self, step: TaskStep, task: Task, **kwargs: Any
    ) -> TaskStepOutput:
        raise NotImplementedError
        # """Run step (async stream)."""
        # return await self._arun_step_stream(step, task)

    def finalize_task(self, task: Task, **kwargs: Any) -> None:
        """Finalize task, after all the steps are completed."""
        # add new messages to memory
        task.memory.put_messages(task.extra_state["new_memory"].get_all())
        # reset new memory
        task.extra_state["new_memory"].reset()

from_tools classmethod #

from_tools(tools: Optional[Sequence[BaseTool]] = None, tool_retriever: Optional[ObjectRetriever[BaseTool]] = None, llm: Optional[LLM] = None, callback_manager: Optional[CallbackManager] = None, verbose: bool = False, **kwargs: Any) -> LLMCompilerAgentWorker

Convenience constructor method from set of of BaseTools (Optional).

Returns:

Name Type Description
LLMCompilerAgentWorker LLMCompilerAgentWorker

the LLMCompilerAgentWorker instance

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
@classmethod
def from_tools(
    cls,
    tools: Optional[Sequence[BaseTool]] = None,
    tool_retriever: Optional[ObjectRetriever[BaseTool]] = None,
    llm: Optional[LLM] = None,
    callback_manager: Optional[CallbackManager] = None,
    verbose: bool = False,
    **kwargs: Any,
) -> "LLMCompilerAgentWorker":
    """Convenience constructor method from set of of BaseTools (Optional).

    Returns:
        LLMCompilerAgentWorker: the LLMCompilerAgentWorker instance

    """
    llm = llm or OpenAI(model=DEFAULT_MODEL_NAME)
    if callback_manager is not None:
        llm.callback_manager = callback_manager
    return cls(
        tools=tools or [],
        tool_retriever=tool_retriever,
        llm=llm,
        callback_manager=callback_manager,
        verbose=verbose,
    )

initialize_step #

initialize_step(task: Task, **kwargs: Any) -> TaskStep

Initialize step from task.

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
def initialize_step(self, task: Task, **kwargs: Any) -> TaskStep:
    """Initialize step from task."""
    sources: List[ToolOutput] = []
    # temporary memory for new messages
    new_memory = ChatMemoryBuffer.from_defaults()

    # put user message in memory
    new_memory.put(ChatMessage(content=task.input, role=MessageRole.USER))

    # initialize task state
    task_state = {
        "sources": sources,
        "new_memory": new_memory,
    }
    task.extra_state.update(task_state)

    return TaskStep(
        task_id=task.task_id,
        step_id=str(uuid.uuid4()),
        input=task.input,
        step_state={"is_replan": False, "contexts": [], "replans": 0},
    )

get_tools #

get_tools(input: str) -> List[AsyncBaseTool]

Get tools.

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
239
240
241
242
def get_tools(self, input: str) -> List[AsyncBaseTool]:
    """Get tools."""
    # return [adapt_to_async_tool(t) for t in self._get_tools(input)]
    return [adapt_to_async_tool(t) for t in self.tools]

arun_llm async #

arun_llm(input: str, previous_context: Optional[str] = None, is_replan: bool = False) -> ChatResponse

Run LLM.

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
async def arun_llm(
    self,
    input: str,
    previous_context: Optional[str] = None,
    is_replan: bool = False,
) -> ChatResponse:
    """Run LLM."""
    if is_replan:
        system_prompt = self.system_prompt_replan
        assert previous_context is not None, "previous_context cannot be None"
        human_prompt = f"Question: {input}\n{previous_context}\n"
    else:
        system_prompt = self.system_prompt
        human_prompt = f"Question: {input}"

    messages = [
        ChatMessage(role=MessageRole.SYSTEM, content=system_prompt),
        ChatMessage(role=MessageRole.USER, content=human_prompt),
    ]

    return await self.llm.achat(messages)

ajoin async #

ajoin(input: str, tasks: Dict[int, LLMCompilerTask], is_final: bool = False) -> JoinerOutput

Join answer using LLM/agent.

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
async def ajoin(
    self,
    input: str,
    tasks: Dict[int, LLMCompilerTask],
    is_final: bool = False,
) -> JoinerOutput:
    """Join answer using LLM/agent."""
    agent_scratchpad = "\n\n"
    agent_scratchpad += "".join(
        [
            task.get_thought_action_observation(
                include_action=True, include_thought=True
            )
            for task in tasks.values()
            if not task.is_join
        ]
    )
    agent_scratchpad = agent_scratchpad.strip()

    output = self.joiner_program(
        query_str=input,
        context_str=agent_scratchpad,
    )
    output = cast(JoinerOutput, output)
    if self.verbose:
        print_text(f"> Thought: {output.thought}\n", color="pink")
        print_text(f"> Answer: {output.answer}\n", color="pink")
    if is_final:
        output.is_replan = False
    return output

run_step #

run_step(step: TaskStep, task: Task, **kwargs: Any) -> TaskStepOutput

Run step.

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
398
399
400
401
@trace_method("run_step")
def run_step(self, step: TaskStep, task: Task, **kwargs: Any) -> TaskStepOutput:
    """Run step."""
    return asyncio.run(self.arun_step(step=step, task=task, **kwargs))

arun_step async #

arun_step(step: TaskStep, task: Task, **kwargs: Any) -> TaskStepOutput

Run step (async).

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
403
404
405
406
407
408
@trace_method("run_step")
async def arun_step(
    self, step: TaskStep, task: Task, **kwargs: Any
) -> TaskStepOutput:
    """Run step (async)."""
    return await self._arun_step(step, task)

stream_step #

stream_step(step: TaskStep, task: Task, **kwargs: Any) -> TaskStepOutput

Run step (stream).

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
410
411
412
413
414
415
@trace_method("run_step")
def stream_step(self, step: TaskStep, task: Task, **kwargs: Any) -> TaskStepOutput:
    """Run step (stream)."""
    # # TODO: figure out if we need a different type for TaskStepOutput
    # return self._run_step_stream(step, task)
    raise NotImplementedError

finalize_task #

finalize_task(task: Task, **kwargs: Any) -> None

Finalize task, after all the steps are completed.

Source code in llama-index-integrations/agent/llama-index-agent-llm-compiler/llama_index/agent/llm_compiler/step.py
425
426
427
428
429
430
def finalize_task(self, task: Task, **kwargs: Any) -> None:
    """Finalize task, after all the steps are completed."""
    # add new messages to memory
    task.memory.put_messages(task.extra_state["new_memory"].get_all())
    # reset new memory
    task.extra_state["new_memory"].reset()