Perplexica is an open-source project that uses large language models (LLMs) and AI capabilities to provide a privacy-preserving alternative to proprietary AI search engines like Perplexity AI.
Perplexica says it uses LMMs combined with machine learning algorithms such as similarity search and embeddings to refine search results and provide consistent answers with sources.
Perplexica's architecture consists of several components: a web-based user interface, agents and chains for predicting next steps, SearXNG for web search, LLMs for understanding content and writing answers, and embedding models for re-ranking search results.
The process works as follows: The user request is sent to the backend server, where the search chain is triggered. It determines whether a web search is required. If so, a search query is sent to SearXNG in normal mode.
The results are converted into embeddings and undergo a similarity search to find the most relevant sources. These sources are passed to the response generator, which does exactly that and sends them to the user interface. The language models cite the sources.
Perplexica has two main modes: The "Copilot Mode" (still under development) aims to find more relevant web sources by generating multiple queries based on the user's search. The "Normal Mode" processes the query and performs a web search.
Video: Perplexica
In addition, Perplexica offers six focus modes designed to provide the best possible answers for specific types of questions: These include an "All Mode" for broad web searches, a "Writing Assistant" for writing assignments, an "Academic Search" for scientific research, a "YouTube Search" for videos, a "Wolfram Alpha Search" for calculations and data analysis, and a "Reddit Search" for discussions and opinions.
Perplexica is best installed via Docker, but can also be installed without it. Step-by-step instructions are available in the installation documentation. Perplexica can be set up as an alternative search engine in browsers.
