WizSearch

Langfuse Traces
Question answering on documents

Inspiration

Inspired by the advanced capabilities of AI assistants like Perplexity, we aimed to create Wiz Search, an open-source, customizable, and modular AI assistant. Our goal is to leverage open-source models and tools to develop a robust and intelligent search assistant that can seamlessly integrate internet searches with large language model (LLM) capabilities.

What it does

Wiz Search is a super-smart AI assistant designed to retrieve and synthesize information from the internet. Users can ask questions, and Wiz Search utilizes powerful LLMs to generate accurate and relevant answers, complete with summaries and links to sources. This makes finding information as easy and magical as simply asking.

How we built it

We built Wiz Search using the following components:

LLM: Snowflakes Arctic LLM for natural language understanding and generation.
Embeddings: Snowflakes Arctic Embedding Model to enhance search relevance.
Intelligent Search: Tavily for advanced search capabilities.
Vector Databases: Qdrant for efficient data storage and retrieval.
Observability: Langfuse for monitoring and observability.
UI: Streamlit for creating an interactive and user-friendly interface.

We focused on creating a modular architecture without relying on LangChain, allowing for a more customizable and flexible design.

Challenges we ran into

Unpredictable LLM Output: Ensuring consistency and accuracy in the responses generated by LLMs.
Retrieval Issues: Addressing the challenge of the retrieval process not always selecting the most relevant information, leading to potential hallucinations in responses.
Security: Implementing guardrails to prevent prompt injection and other security vulnerabilities.

Accomplishments that we're proud of

Successfully integrating multiple open-source tools to create a seamless and efficient search assistant.
Developing a modular and customizable workflow that allows for easy adjustments and enhancements.

What we learned

The importance of modular RAG (retrieval-augmented generation) and LLM routing for maintaining controlled and efficient agentic workflows.
Effective prompting techniques for open-source models to optimize performance.

What's next for Wiz Search

In the future, we aim to enhance Wiz Search by incorporating more controlled agentic capabilities, such as running code and generating images. We also plan to develop a minimalistic, voice-interactive platform that responds in a human-like voice and displays only the most relevant visual information, such as images and graphs. Technically, we will integrate RAGAS for evaluation and Giskard for testing. This will further streamline the user experience and broaden the range of tasks Wiz Search can perform.