[BUG] Knowledge doesn't read from the knowledge source using StringSource

[chadsly]

Description

I've attempted running the example "knowledge" code that is in the docs: https://docs.crewai.com/concepts/knowledge
The first example uses the StringSource. Maybe after I can get this one to work, I'll move on to more complicated options.
I used the code as is (except changing the llm) and run it as a simple python script. I've also rewritten it to run it as a crewai input.

Steps to Reproduce

1. With the python code below, simply run python3 main.py

Expected behavior

1. I expected the output of the question to give the information about the user ("John") that corresponded to the knowledge source. In this case, it should have told me that John was 30 and living in San Fransisco.

Screenshots/Code snippets

from crewai import Agent, Task, Crew, Process, LLM
from crewai.knowledge.source.string_knowledge_source import StringKnowledgeSource

Create a knowledge source

content = "Users name is John. He is 30 years old and lives in San Francisco."
string_source = StringKnowledgeSource(
content=content,
)

Create an LLM with a temperature of 0 to ensure deterministic outputs

llm = LLM(model="gpt-4o-mini", temperature=0)

llm=LLM(model="ollama/llama3.2:1b", base_url="http://localhost:11434")

Create an agent with the knowledge store

agent = Agent(
role="About User",
goal="You know everything about the user.",
backstory="""You are a master at understanding people and their preferences.""",
verbose=True,
allow_delegation=False,
llm=llm,
)
task = Task(
description="Answer the following questions about the user: {question}",
expected_output="An answer to the question.",
agent=agent,
)

crew = Crew(
agents=[agent],
tasks=[task],
verbose=True,
process=Process.sequential,
knowledge_sources=[string_source], # Enable knowledge by adding the sources here. You can also add more sources to the sources list.
)

result = crew.kickoff(inputs={"question": "What city does John Albertson live in and how old is he?"})

Operating System

Ubuntu 20.04

Python Version

3.10

crewAI Version

0.98.0

crewAI Tools Version

Virtual Environment

Venv

Evidence

The answer told me that John lived in New York City and was 38 years old. In other words, it didn't use the knowledge at all.

Possible Solution

None

Additional context

1. I didn't include the larger crewai CLI example because it seemed to have the same incorrect information. It's just not reading the knowledge source.
2. In case you didn't notice in the code, I'm using ollama with a local model.

[Vidit-Ostwal]

Hi @chadsly, I think there is a issue with the documentation, can you try passing a embedder at the crew itself.
As a general rule, any knowledge source will require a embedder, because that's how they are stored internally.

Meanwhile I will raise a PR for this.

[Vidit-Ostwal]

Can you try this once, I am facing some other issue, that's why not able to test, let me know once whether this solves the entire thing or not.

[chadsly]

I changed the crew to include an embedder...
crew = Crew(
agents=[agent],
tasks=[task],
verbose=True,
process=Process.sequential,
knowledge_sources=[string_source], # Enable knowledge by adding the sources here. You can also add more sources to the sources list.
embedder={
"provider": "ollama",
"config": {
"model": "llama3.2:1b",
}
}
)

It's the same model as the llm. I could have used something else, but I didn't. I also checked to see if I could include a base_url. It didn't seem to make any difference. I'll make this a separate ticket if I ever run into that issue.

[lorenzejay]

@chadsly use an embedding model: with ollama download:

granite-embedding:30m
or
all-minilm:latest

Embedding models are supported here instead of a text completion model

[chadsly]

@lorenzejay , well now I'm stuck in an awkward situation where I can't change the embedding model, which sounds like an environment issue.

In the crewai projec that I'm working with I've tried to
crewai reset-memories -a
but because I'm using ollama it seems to get confused with

An unexpected error occurred: Please provide an OpenAI API key. You can get one at https://platform.openai.com/account/api-keys

Clearly I'm not using any OpenAI keys.

So I simply started a new environment to work in. This worked. I did not know that I couldn't embedding models. Is that a crewai limitation? a library limitation? or just how embedding works?

Summary: My original request has been answered.

[lorenzejay]

Please upgrade to the latest version. That should've been fixed. And for anything of storing to a vector db, we require embedding models to turn text into a representation that can be semantically represented in a vector db. Just how it works compared to a chat completion.