-
Notifications
You must be signed in to change notification settings - Fork 159
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
feat: Add deep research use case (Python) (#482)
--------- Co-authored-by: Marcus Schiesser <[email protected]>
- Loading branch information
Showing
17 changed files
with
928 additions
and
22 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,5 @@ | ||
| --- | ||
| "create-llama": patch | ||
| --- | ||
|
|
||
| Add deep research over own documents use case (Python) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.
Oops, something went wrong.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
47 changes: 47 additions & 0 deletions
47
templates/components/agents/python/deep_research/README-template.md
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,47 @@ | ||
| This is a [LlamaIndex](https://www.llamaindex.ai/) multi-agents project using [Workflows](https://docs.llamaindex.ai/en/stable/understanding/workflows/). | ||
|
|
||
| ## Getting Started | ||
|
|
||
| First, setup the environment with poetry: | ||
|
|
||
| > **_Note:_** This step is not needed if you are using the dev-container. | ||
| ```shell | ||
| poetry install | ||
| ``` | ||
|
|
||
| Then check the parameters that have been pre-configured in the `.env` file in this directory. (E.g. you might need to configure an `OPENAI_API_KEY` if you're using OpenAI as model provider). | ||
| Second, generate the embeddings of the documents in the `./data` directory: | ||
|
|
||
| ```shell | ||
| poetry run generate | ||
| ``` | ||
|
|
||
| Third, run the development server: | ||
|
|
||
| ```shell | ||
| poetry run dev | ||
| ``` | ||
|
|
||
| ## Use Case: Deep Research over own documents | ||
|
|
||
| The workflow performs deep research by retrieving and analyzing documents from the [data](./data) directory from multiple perspectives. The project includes a sample PDF about AI investment in 2024 to help you get started. You can also add your own documents by placing them in the data directory and running the generate script again to index them. | ||
|
|
||
| After starting the server, go to [http://localhost:8000](http://localhost:8000) and send a message to the agent to write a blog post. | ||
| E.g: "AI investment in 2024" | ||
|
|
||
| To update the workflow, you can edit the [deep_research.py](./app/workflows/deep_research.py) file. | ||
|
|
||
| By default, the workflow retrieves 10 results from your documents. To customize the amount of information covered in the answer, you can adjust the `TOP_K` environment variable in the `.env` file. A higher value will retrieve more results from your documents, potentially providing more comprehensive answers. | ||
|
|
||
| ## Deployments | ||
|
|
||
| For production deployments, check the [DEPLOY.md](DEPLOY.md) file. | ||
|
|
||
| ## Learn More | ||
|
|
||
| To learn more about LlamaIndex, take a look at the following resources: | ||
|
|
||
| - [LlamaIndex Documentation](https://docs.llamaindex.ai) - learn about LlamaIndex. | ||
| - [Workflows Introduction](https://docs.llamaindex.ai/en/stable/understanding/workflows/) - learn about LlamaIndex workflows. | ||
| You can check out [the LlamaIndex GitHub repository](https://github.com/run-llama/llama_index) - your feedback and contributions are welcome! |
3 changes: 3 additions & 0 deletions
3
templates/components/agents/python/deep_research/app/workflows/__init__.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,3 @@ | ||
| from .deep_research import create_workflow | ||
|
|
||
| __all__ = ["create_workflow"] |
158 changes: 158 additions & 0 deletions
158
templates/components/agents/python/deep_research/app/workflows/agents.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,158 @@ | ||
| from typing import List, Literal, Optional | ||
|
|
||
| from llama_index.core.base.llms.types import ( | ||
| CompletionResponse, | ||
| CompletionResponseAsyncGen, | ||
| ) | ||
| from llama_index.core.memory.simple_composable_memory import SimpleComposableMemory | ||
| from llama_index.core.prompts import PromptTemplate | ||
| from llama_index.core.schema import MetadataMode, Node, NodeWithScore | ||
| from llama_index.core.settings import Settings | ||
| from pydantic import BaseModel, Field | ||
|
|
||
|
|
||
| class AnalysisDecision(BaseModel): | ||
| decision: Literal["research", "write", "cancel"] = Field( | ||
| description="Whether to continue research, write a report, or cancel the research after several retries" | ||
| ) | ||
| research_questions: Optional[List[str]] = Field( | ||
| description="Questions to research if continuing research. Maximum 3 questions. Set to null or empty if writing a report.", | ||
| default_factory=list, | ||
| ) | ||
| cancel_reason: Optional[str] = Field( | ||
| description="The reason for cancellation if the decision is to cancel research.", | ||
| default=None, | ||
| ) | ||
|
|
||
|
|
||
| async def plan_research( | ||
| memory: SimpleComposableMemory, | ||
| context_nodes: List[Node], | ||
| user_request: str, | ||
| ) -> AnalysisDecision: | ||
| analyze_prompt = PromptTemplate( | ||
| """ | ||
| You are a professor who is guiding a researcher to research a specific request/problem. | ||
| Your task is to decide on a research plan for the researcher. | ||
| The possible actions are: | ||
| + Provide a list of questions for the researcher to investigate, with the purpose of clarifying the request. | ||
| + Write a report if the researcher has already gathered enough research on the topic and can resolve the initial request. | ||
| + Cancel the research if most of the answers from researchers indicate there is insufficient information to research the request. Do not attempt more than 3 research iterations or too many questions. | ||
| The workflow should be: | ||
| + Always begin by providing some initial questions for the researcher to investigate. | ||
| + Analyze the provided answers against the initial topic/request. If the answers are insufficient to resolve the initial request, provide additional questions for the researcher to investigate. | ||
| + If the answers are sufficient to resolve the initial request, instruct the researcher to write a report. | ||
| <User request> | ||
| {user_request} | ||
| </User request> | ||
| <Collected information> | ||
| {context_str} | ||
| </Collected information> | ||
| <Conversation context> | ||
| {conversation_context} | ||
| </Conversation context> | ||
| """ | ||
| ) | ||
| conversation_context = "\n".join( | ||
| [f"{message.role}: {message.content}" for message in memory.get_all()] | ||
| ) | ||
| context_str = "\n".join( | ||
| [node.get_content(metadata_mode=MetadataMode.LLM) for node in context_nodes] | ||
| ) | ||
| res = await Settings.llm.astructured_predict( | ||
| output_cls=AnalysisDecision, | ||
| prompt=analyze_prompt, | ||
| user_request=user_request, | ||
| context_str=context_str, | ||
| conversation_context=conversation_context, | ||
| ) | ||
| return res | ||
|
|
||
|
|
||
| async def research( | ||
| question: str, | ||
| context_nodes: List[NodeWithScore], | ||
| ) -> str: | ||
| prompt = """ | ||
| You are a researcher who is in the process of answering the question. | ||
| The purpose is to answer the question based on the collected information, without using prior knowledge or making up any new information. | ||
| Always add citations to the sentence/point/paragraph using the id of the provided content. | ||
| The citation should follow this format: [citation:id]() where id is the id of the content. | ||
| E.g: | ||
| If we have a context like this: | ||
| <Citation id='abc-xyz'> | ||
| Baby llama is called cria | ||
| </Citation id='abc-xyz'> | ||
| And your answer uses the content, then the citation should be: | ||
| - Baby llama is called cria [citation:abc-xyz]() | ||
| Here is the provided context for the question: | ||
| <Collected information> | ||
| {context_str} | ||
| </Collected information>` | ||
| No prior knowledge, just use the provided context to answer the question: {question} | ||
| """ | ||
| context_str = "\n".join( | ||
| [_get_text_node_content_for_citation(node) for node in context_nodes] | ||
| ) | ||
| res = await Settings.llm.acomplete( | ||
| prompt=prompt.format(question=question, context_str=context_str), | ||
| ) | ||
| return res.text | ||
|
|
||
|
|
||
| async def write_report( | ||
| memory: SimpleComposableMemory, | ||
| user_request: str, | ||
| stream: bool = False, | ||
| ) -> CompletionResponse | CompletionResponseAsyncGen: | ||
| report_prompt = """ | ||
| You are a researcher writing a report based on a user request and the research context. | ||
| You have researched various perspectives related to the user request. | ||
| The report should provide a comprehensive outline covering all important points from the researched perspectives. | ||
| Create a well-structured outline for the research report that covers all the answers. | ||
| # IMPORTANT when writing in markdown format: | ||
| + Use tables or figures where appropriate to enhance presentation. | ||
| + Preserve all citation syntax (the `[citation:id]()` parts in the provided context). Keep these citations in the final report - no separate reference section is needed. | ||
| + Do not add links, a table of contents, or a references section to the report. | ||
| <User request> | ||
| {user_request} | ||
| </User request> | ||
| <Research context> | ||
| {research_context} | ||
| </Research context> | ||
| Now, write a report addressing the user request based on the research provided following the format and guidelines above. | ||
| """ | ||
| research_context = "\n".join( | ||
| [f"{message.role}: {message.content}" for message in memory.get_all()] | ||
| ) | ||
|
|
||
| llm_complete_func = ( | ||
| Settings.llm.astream_complete if stream else Settings.llm.acomplete | ||
| ) | ||
|
|
||
| res = await llm_complete_func( | ||
| prompt=report_prompt.format( | ||
| user_request=user_request, | ||
| research_context=research_context, | ||
| ), | ||
| ) | ||
| return res | ||
|
|
||
|
|
||
| def _get_text_node_content_for_citation(node: NodeWithScore) -> str: | ||
| """ | ||
| Construct node content for LLM with citation flag. | ||
| """ | ||
| node_id = node.node.node_id | ||
| content = f"<Citation id='{node_id}'>\n{node.get_content(metadata_mode=MetadataMode.LLM)}</Citation id='{node_id}'>" | ||
| return content |
Oops, something went wrong.