Note
Pixlie is currently in beta. We value your feedback as we progress towards a stable release.
AI powered knowledge graphs for semantically accurate insights. From online search to private document search to semantic search in your apps.
Pixlie is an autonomous knowledge construction system that transforms how you interact with information. Beyond simple data retrieval, Pixlie builds rich semantic understanding through dynamic knowledge graphs. At its core, Pixlie employs a multi-agent AI architecture that:
- Understands your objective - Start with a simple prompt or question
- Autonomously discovers relevant information - Our intelligent agents extract, connect, and contextualize data
- Constructs a living knowledge graph - Entities and relationships are explicitly modeled, not just vectorized
- Self-improves as it learns - Internal prompts continuously optimize based on new discoveries
Unlike vector databases that flatten semantic meaning into numerical proximities, Pixlie preserves the true structure of information - people connect to organizations, events happen at specific times, concepts relate to each other through defined relationships - creating a foundation for genuinely intelligent insights.
Vector databases are good for storing and querying semantic data, but they don't model the underlying data accurately. Pixlie uses a sophisticated hybrid approach:
- Multi-model intelligence - We leverage both large language models and specialized smaller models tailored to specific tasks (entity recognition, classification, relationship extraction)
- Task-specific agents - Different components of our system use the most appropriate model for each job, balancing accuracy with efficiency
- Granular semantic modeling - Each entity (person, place, date, event) is stored as a distinct node in the graph with explicit relationships to other entities
- Adaptive processing pipeline - The system selects the optimal processing approach based on the content and objective
This architecture makes our approach more accurate for complex information needs while remaining computationally efficient and cost effective.
We're committed to true data privacy. Pixlie is completely self-hosted with no data leaving your environment - a crucial advantage in today's regulatory landscape.
We believe in transparency and community-driven development. Our core offering is open source with optional paid features for enterprise needs and we are building in public.
- Objective driven projects to create knowledge graphs ("Track companies on Indian stock exchanges")
- Privacy-first architecture - completely self-hosted with no data leaving your environment
- Use Anthropic Claude (other models' support coming soon) with your own API keys
- Use Brave Search API for web search with your own API keys
- Built-in simple web crawler
- Runs anywhere - on your laptop or on the cloud
- Team collaboration (paid feature)
- Entity extraction - identify people, places, dates, events, etc. (some paid features)
- Explore your knowledge graph (coming soon)
- Open source with transparent development
Watch our latest product update on YouTube:
Choose your objective - what do you want to discover with Pixlie?:
Information will instantly be populated. Let Pixlie search a while and see...
A list of links discovered based on your objective:
A list of domains (as many links may come from one domain):
The content extracted from these links:
Search through the results for specifics:
We're actively developing Pixlie with these exciting features on the horizon:
- Knowledge Graph Explorer - Interactive visual interface to browse, explore, and edit your knowledge graph
- Self-Correcting Prompt System - AI that automatically refines its prompts as it learns from new information
- Multi-Agent Architecture - Specialized agents working together (crawler, classifier, analyzer) to build comprehensive knowledge graphs
- Notifications of new matches on monitored websites
- Desktop Application - Native experience across platforms
- Advanced Document Support - Ingest CSV, JSON, Markdown, and more file formats beyond web content
At the moment Pixlie can be used on your laptop for personal web research (desktop app coming soon!). Pixlie is open source, with the option of enterprise licensing for those who want to utilize Pixlie's search and knowledge graph within their product.
Pixlie is under active development. Our work is done in public, please star this repository, it means a lot to us.
If you want to develop on Pixlie, please see the DEVELOP.md file.
If you want to use Pixlie, please see the USE.md file.
Pixlie has a REST API that you can use to interact with the graph. We use Bruno to document the API.
You can find the API spec in the rest_api directory.
We are happily taking contributions! If you are interested in contributing to Pixlie, please join our Discord Channel or reach out to us to set up a call. Share where you want to contribute. We have a lot of to dos stuck in our head that we are happy to put on our project board if we know someone is interested in taking it on.
If you want to be notified when Pixlie is ready for use, please subscribe to our insights newsletter.
Join the discussion or get support on Discord.
- Pixlie is licensed under the GNU General Public License version 3.0
- See the LICENSE file for details
Yes, Pixlie is an alternative to using vector databases. Vector databases are good for storing and querying semantic data, but they do not model the underlying data accurately. In Pixlie, we leverage both LLMs and specialized smaller models tailored to specific tasks (entity recognition, classification, relationship extraction). Each individual entity, such as a person, place, date, event, etc., is stored separately in the graph, along with its relationships to other entities. This makes our graph based approach better where accuracy is important.
Link Nodes are the URLs we discover using Brave & AI based on the Objective. When each Link Node is processed, we store domains for these Links in our database as separate nodes - one node for every domain.
We connect Domain Nodes to their Link Nodes in the following way:
Domain —OwnerOf—> Link(s)
Link —BelongsTo—> a Domain
As a part of the processing of Link Nodes, we then fetch their raw HTML content. Once we successfully fetch the content, we create a WebPage Node to store that content and then connect it to the Link Node:
Link —PathOf—> WebPage
and vice-versa
WebPage —ContentOf—> Link