The way to Construct an AI Journal with LlamaIndex

This put up will share how one can construct an AI journal with the LlamaIndex. We’ll cowl one important operate of this AI journal: asking for recommendation. We’ll begin with essentially the most fundamental implementation and iterate from there. We will see important enhancements for this operate once we apply design patterns like Agentic Rag and multi-agent workflow.

Yow will discover the supply code of this AI Journal in my GitHub repo here. And about who I am.

Overview of AI Journal

I wish to construct my rules by following Ray Dalio’s follow. An AI journal will assist me to self-reflect, monitor my enchancment, and even give me recommendation. The general operate of such an AI journal appears like this:

In the present day, we are going to solely cowl the implementation of the seek-advise stream, which is represented by a number of purple cycles within the above diagram.

Easiest Kind: LLM with Massive Context

In essentially the most easy implementation, we will go all of the related content material into the context and connect the query we wish to ask. We will try this in Llamaindex with just a few traces of code.

import pymupdf
from llama_index.llms.openai import OpenAI

path_to_pdf_book = './path/to/pdf/e book.pdf'
def load_book_content():
    textual content = ""
    with pymupdf.open(path_to_pdf_book) as pdf:
        for web page in pdf:
            textual content += str(web page.get_text().encode("utf8", errors='ignore'))
    return textual content

system_prompt_template = """You might be an AI assistant that gives considerate, sensible, and *deeply customized* ideas by combining:
- The consumer's private profile and rules
- Insights retrieved from *Ideas* by Ray Dalio
E-book Content material: 
```
{book_content}
```
Consumer profile:
```
{user_profile}
```
Consumer's query:
```
{user_question}
```
"""

def get_system_prompt(book_content: str, user_profile: str, user_question: str):
    system_prompt = system_prompt_template.format(
        book_content=book_content,
        user_profile=user_profile,
        user_question=user_question
    )
    return system_prompt

def chat():
    llm = get_openai_llm()
    user_profile = enter(">>Inform me about your self: ")
    user_question = enter(">>What do you wish to ask: ")
    user_profile = user_profile.strip()
    book_content = load_book_summary()
    response = llm.full(immediate=get_system_prompt(book_content, user_profile, user_question))
    return response

This strategy has downsides:

• Low Precision: Loading all of the e book context may immediate LLM to lose deal with the consumer’s query.
• Excessive Value: Sending over significant-sized content material in each LLM name means excessive value and poor efficiency.

With this strategy, if you happen to go the entire content material of Ray Dalio’s Ideas e book, responses to questions like “The way to deal with stress?” grow to be very common. Such responses with out regarding my query made me really feel that the AI was not listening to me. Though it covers many vital ideas like embracing actuality, the 5-step course of to get what you need, and being radically open-minded. I like the recommendation I bought to be extra focused to the query I raised. Let’s see how we will enhance it with RAG.

Enhanced Kind: Agentic RAG

So, what’s Agentic RAG? Agentic RAG is combining dynamic decision-making and knowledge retrieval. In our AI journal, the Agentic RAG stream appears like this:

• Query Analysis: Poorly framed questions result in poor question outcomes. The agent will consider the consumer’s question and make clear the questions if the Agent believes it’s mandatory.
• Query Re-write: Rewrite the consumer enquiry to challenge it to the listed content material within the semantic area. I discovered these steps important for enhancing the precision in the course of the retrieval. Let’s say in case your data base is Q/A pair and you’re indexing the questions half to seek for solutions. Rewriting the consumer’s question assertion to a correct query will assist you to discover essentially the most related content material.
• Question Vector Index: Many parameters might be tuned when constructing such an index, together with chunk dimension, overlap, or a unique index sort. For simplicity, we’re utilizing VectorStoreIndex right here, which has a default chunking technique.
• Filter & Artificial: As an alternative of a posh re-ranking course of, I explicitly instruct LLM to filter and discover related content material within the immediate. I see LLM selecting up essentially the most related content material, though typically it has a decrease similarity rating than others.

With this Agentic RAG, you’ll be able to retrieve extremely related content material to the consumer’s questions, producing extra focused recommendation.

Let’s study the implementation. With the LlamaIndex SDK, creating and persisting an index in your native listing is easy.

from llama_index.core import Doc, VectorStoreIndex, StorageContext, load_index_from_storage

Settings.embed_model = OpenAIEmbedding(api_key="ak-xxxx")
PERSISTED_INDEX_PATH = "/path/to/the/listing/persist/index/domestically"

def create_index(content material: str):
    paperwork = [Document(text=content)]
    vector_index = VectorStoreIndex.from_documents(paperwork)
    vector_index.storage_context.persist(persist_dir=PERSISTED_INDEX_PATH)

def load_index():
    storage_context = StorageContext.from_defaults(persist_dir=PERSISTED_INDEX_PATH)
    index = load_index_from_storage(storage_context)
    return index

As soon as we now have an index, we will create a question engine on high of that. The question engine is a robust abstraction that permits you to alter the parameters in the course of the question(e.g., TOP Okay) and the synthesis behaviour after the content material retrieval. In my implementation, I overwrite the response_mode NO_TEXT as a result of the agent will course of the e book content material returned by the operate name and synthesize the ultimate end result. Having the question engine to synthesize the end result earlier than passing it to the agent could be redundant.

from llama_index.core.indices.vector_store import VectorIndexRetriever
from llama_index.core.query_engine import RetrieverQueryEngine
from llama_index.core.response_synthesizers import ResponseMode
from llama_index.core import  VectorStoreIndex, get_response_synthesizer

def _create_query_engine_from_index(index: VectorStoreIndex):
    # configure retriever
    retriever = VectorIndexRetriever(
        index=index,
        similarity_top_k=TOP_K,
    )
    # return the unique content material with out utilizing LLM to synthesizer. For later analysis.
    response_synthesizer = get_response_synthesizer(response_mode=ResponseMode.NO_TEXT)
    # assemble question engine
    query_engine = RetrieverQueryEngine(
        retriever=retriever,
        response_synthesizer=response_synthesizer
    )
    return query_engine

The immediate appears like the next:

You might be an assistant that helps reframe consumer questions into clear, concept-driven statements that match 
the model and matters of Ideas by Ray Dalio, and carry out search for precept e book for related content material. 

Background:
Ideas teaches structured enthusiastic about life and work choices.
The important thing concepts are:
* Radical reality and radical transparency
* Choice-making frameworks
* Embracing errors as studying

Activity:
- Activity 1: Make clear the consumer's query if wanted. Ask follow-up questions to make sure you perceive the consumer's intent.
- Activity 2: Rewrite a consumer’s query into a press release that will match how Ray Dalio frames concepts in Ideas. Use formal, logical, impartial tone.
- Activity 3: Search for precept e book with given re-wrote statements. It is best to present no less than {REWRITE_FACTOR} rewrote variations.
- Activity 4: Discover essentially the most related from the e book content material as your fina solutions.

Lastly, we will construct the agent with these capabilities outlined.

def get_principle_rag_agent():
    index = load_persisted_index()
    query_engine = _create_query_engine_from_index(index)

    def look_up_principle_book(original_question: str, rewrote_statement: Listing[str]) -> Listing[str]:
        end result = []
        for q in rewrote_statement:
            response = query_engine.question(q)
            content material = [n.get_content() for n in response.source_nodes]
            end result.lengthen(content material)
        return end result

    def clarify_question(original_question: str, your_questions_to_user: Listing[str]) -> str:
        """
        Make clear the consumer's query if wanted. Ask follow-up questions to make sure you perceive the consumer's intent.
        """
        response = ""
        for q in your_questions_to_user:
            print(f"Query: {q}")
            r = enter("Response:")
            response += f"Query: {q}nResponse: {r}n"
        return response

    instruments = [
        FunctionTool.from_defaults(
            fn=look_up_principle_book,
            name="look_up_principle_book",
            description="Look up principle book with re-wrote queries. Getting the suggestions from the Principle book by Ray Dalio"),
        FunctionTool.from_defaults(
            fn=clarify_question,
            name="clarify_question",
            description="Clarify the user's question if needed. Ask follow-up questions to ensure you understand the user's intent.",
        )
    ]

    agent = FunctionAgent(
        identify="principle_reference_loader",
        description="You're a useful agent will primarily based on consumer's query and search for essentially the most related content material in precept e book.n",
        system_prompt=QUESTION_REWRITE_PROMPT,
        instruments=instruments,
    )
    return agent

rag_agent = get_principle_rag_agent()
response = await agent.run(chat_history=chat_history)

There are just a few observations I had in the course of the implementations:

• One attention-grabbing reality I discovered is that offering a non-used parameter, original_question , within the operate signature helps. I discovered that once I don’t have such a parameter, LLM typically doesn’t observe the rewrite instruction and passes the unique query in rewrote_statement the parameter. Having original_question parameters in some way emphasizes the rewriting mission to LLM.
• Completely different LLMs behave fairly in a different way given the identical immediate. I discovered DeepSeek V3 far more reluctant to set off operate calls than different mannequin suppliers. This doesn’t essentially imply it’s not usable. If a useful name must be initiated 90% of the time, it must be a part of the workflow as an alternative of being registered as a operate name. Additionally, in comparison with OpenAI’s fashions, I discovered Gemini good at citing the supply of the e book when it synthesizes the outcomes.
• The extra content material you load into the context window, the extra inference functionality the mannequin wants. A smaller mannequin with much less inference energy is extra more likely to get misplaced within the giant context offered.

Nonetheless, to finish the seek-advice operate, you’ll want a number of Brokers working collectively as an alternative of a single Agent. Let’s speak about how one can chain your Brokers collectively into workflows.

Remaining Kind: Agent Workflow

Earlier than we begin, I like to recommend this text by Anthropic, Building Effective Agents. The one-liner abstract of the articles is that you need to at all times prioritise constructing a workflow as an alternative of a dynamic agent when potential. In LlamaIndex, you are able to do each. It permits you to create an agent workflow with extra automated routing or a customized workflow with extra express management of the transition of steps. I’ll present an instance of each implementations.

Let’s check out how one can construct a dynamic workflow. Here’s a code instance.

interviewer = FunctionAgent(
        identify="interviewer",
        description="Helpful agent to make clear consumer's questions",
        system_prompt=_intervierw_prompt,
        can_handoff_to = ["retriver"]
        instruments=instruments
)
interviewer = FunctionAgent(
        identify="retriever",
        description="Helpful agent to retrive precept e book's content material.",
        system_prompt=_retriver_prompt,
        can_handoff_to = ["advisor"]
        instruments=instruments
)
advisor = FunctionAgent(
        identify="advisor",
        description="Helpful agent to advise consumer.",
        system_prompt=_advisor_prompt,
        can_handoff_to = []
        instruments=instruments
)
workflow = AgentWorkflow(
        brokers=[interviewer, advisor, retriever],
        root_agent="interviewer",
    )
handler = await workflow.run(user_msg="The way to deal with stress?")

It’s dynamic as a result of the Agent transition relies on the operate name of the LLM mannequin. Underlying, LlamaIndex workflow offers agent descriptions as capabilities for LLM fashions. When the LLM mannequin triggers such “Agent Operate Name”, LlamaIndex will path to your subsequent corresponding agent for the following step processing. Your earlier agent’s output has been added to the workflow inner state, and your following agent will decide up the state as a part of the context of their name to the LLM mannequin. You additionally leverage state and reminiscence parts to handle the workflow’s inner state or load exterior knowledge(reference the doc here).

Nonetheless, as I’ve advised, you’ll be able to explicitly management the steps in your workflow to realize extra management. With LlamaIndex, it may be achieved by extending the workflow object. For instance:

class ReferenceRetrivalEvent(Occasion):
    query: str

class Recommendation(Occasion):
    rules: Listing[str]
    profile: dict
    query: str
    book_content: str

class AdviceWorkFlow(Workflow):
    def __init__(self, verbose: bool = False, session_id: str = None):
        state = get_workflow_state(session_id)
        self.rules = state.load_principle_from_cases()
        self.profile = state.load_profile()
        self.verbose = verbose
        tremendous().__init__(timeout=None, verbose=verbose)

    @step
    async def interview(self, ctx: Context,
                        ev: StartEvent) -> ReferenceRetrivalEvent:
        # Step 1: Interviewer agent asks inquiries to the consumer
        interviewer = get_interviewer_agent()
        query = await _run_agent(interviewer, query=ev.user_msg, verbose=self.verbose)

        return ReferenceRetrivalEvent(query=query)

    @step
    async def retrieve(self, ctx: Context, ev: ReferenceRetrivalEvent) -> Recommendation:
        # Step 2: RAG agent retrieves related content material from the e book
        rag_agent = get_principle_rag_agent()
        book_content = await _run_agent(rag_agent, query=ev.query, verbose=self.verbose)
        return Recommendation(rules=self.rules, profile=self.profile,
                      query=ev.query, book_content=book_content)

    @step
    async def recommendation(self, ctx: Context, ev: Recommendation) -> StopEvent:
        # Step 3: Adviser agent offers recommendation primarily based on the consumer's profile, rules, and e book content material
        advisor = get_adviser_agent(ev.profile, ev.rules, ev.book_content)
        advise = await _run_agent(advisor, query=ev.query, verbose=self.verbose)
        return StopEvent(end result=advise)

The precise occasion sort’s return controls the workflow’s step transition. For example, retrieve step returns an Recommendation occasion that can set off the execution of the recommendation step. You may also leverage the Recommendation occasion to go the required data you want.

Through the implementation, if you’re irritated by having to begin over the workflow to debug some steps within the center, the context object is important while you wish to failover the workflow execution. You may retailer your state in a serialised format and get better your workflow by unserialising it to a context object. Your workflow will proceed executing primarily based on the state as an alternative of beginning over.

workflow = AgentWorkflow(
    brokers=[interviewer, advisor, retriever],
    root_agent="interviewer",
)
strive:
    handler = w.run()
    end result = await handler
besides Exception as e:
    print(f"Error throughout preliminary run: {e}")
    await fail_over()
    # Elective, serialised and save the contexct for debugging 
    ctx_dict = ctx.to_dict(serializer=JsonSerializer())
    json_dump_and_save(ctx_dict)
    # Resume from the identical context
    ctx_dict = load_failed_dict()
    restored_ctx = Context.from_dict(workflow, ctx_dict,serializer=JsonSerializer())
    handler = w.run(ctx=handler.ctx)
    end result = await handler

Abstract

On this put up, we now have mentioned how one can use LlamaIndex to implement an AI journal’s core operate. The important thing studying consists of:

• Utilizing Agentic RAG to leverage LLM functionality to dynamically rewrite the unique question and synthesis end result.
• Use a Custom-made Workflow to realize extra express management over step transitions. Construct dynamic brokers when mandatory.

The bitterce code of this AI journal is in my GitHub repo here. I hope you take pleasure in this text and this small app I constructed. Cheers!

The put up How to Build an AI Journal with LlamaIndex appeared first on Towards Data Science.