Isaacus
then open sourced the code
Last verified:
What is Isaacus?
Isaacus Tabular Review is an open-source AI-powered tabular review application designed for legal professionals that transforms unstructured documents into structured labels and entities without using any generative models. The tool builds hallucination-free knowledge graphs from legal documents using specialized encoder-based models (Kanon 2 Enricher, Kanon 2 Embedder, and Kanon Answer Extractor), offering superior accuracy, lower latency, and zero generative hallucinations compared to competitors like Harvey and Legora.
Key features include automatic entity extraction (persons, organizations, locations, dates, roles), span-level classification using semantic search with configurable similarity thresholds, entity linking and relationship extraction, hierarchical document segmentation, backlinks and navigation panels for exploring document structures, custom column queries using natural language, and an interactive web interface for uploading documents and adding columns. The application uses Qdrant vector database for storing embeddings and implements token-level document annotation with statistically informative confidence scores.
This tool is specifically designed for lawyers and legal professionals who need to compare key terms across many documents during legal review workflows, legal tech developers looking to build or extend tabular review capabilities, and organizations seeking a cost-effective open-source alternative to expensive commercial legal AI tools. The guide is free to use, adapt, extend, and commercialize.
Isaacus pricing
Pricing model: Freemium
The Isaacus Tabular Review guide and code are completely open source and free to use, adapt, extend, and commercialize. An Isaacus account with valid API key is required to access the Kanon 2 Enricher, Kanon 2 Embedder, and Kanon Answer Extractor models via API. The webpage does not specify pricing details for the Isaacus API itself - users must obtain an API key through the API quickstart guide. No paid plans or subscription tiers are mentioned for the tabular review application itself.
Isaacus pros
- Zero generative hallucinations by using only encoder-based models
- Open source and free to use, adapt, extend, and commercialize
- More advanced functionality than Harvey and Legora tabular review tools
- Fraction of the cost compared to commercial competitors
- Greater accuracy than generative model-based alternatives
- Lower latency than generative model pipelines
- Token-level document annotation with confidence scores
- Automatic knowledge graph construction in milliseconds
- Hierarchical document segmentation using ILDGS schema
- Span-level classification with configurable similarity threshold
- Natural language queries for custom column creation
- Entity linking creates relationships on the fly without pre-definition
- Backlinks and panels enable navigation across document structures
- Built-in OCR support for document processing
- In-memory Qdrant vector database for fast embedding storage
- FastAPI server with async background index building
- Extracts persons, organizations, locations, dates, roles automatically
- Cross-references and document relationships tracking
- Percentage-based similarity scoring for classification results
- Parent span preference deduplication for better context
Isaacus cons
- Requires Isaacus account and API key to use
- Requires programming knowledge to implement from scratch
- Uses in-memory Qdrant which doesn't persist data across restarts
- No built-in user authentication or multi-user support
- Tutorial requires setting up FastAPI server manually
- Limited to legal documents optimized for ILDGS schema
- No pre-built GUI - requires cloning Isaacus Cookbook client
- Embedding model task parameter must be set correctly
- Similarity threshold of 0.4 may need tuning for different tasks
- Python dependencies must be installed manually
Frequently asked questions about Isaacus
What is Isaacus Tabular Review?
Isaacus Tabular Review is a state-of-the-art open-source tabular review application that transforms unstructured legal documents into structured labels and entities without using any generative models. It builds hallucination-free knowledge graphs using specialized encoder-based models (Kanon 2 Enricher, Kanon 2 Embedder, and Kanon Answer Extractor), offering greater accuracy, lower latency, and zero generative hallucinations compared to competitors like Harvey and Legora.
How does it achieve hallucination-free results?
The tool achieves hallucination-free results by not using any generative models at all. Instead, it relies entirely on specialized encoder-based models (Kanon 2 Enricher, Kanon 2 Embedder, and Kanon Answer Extractor) that directly annotate documents at the token level. Encoders are fundamentally different from generative models - they cannot generate text that isn't grounded in the source document, making hallucinations mathematically impossible in this architecture.
What models does Isaacus Tabular Review use?
The application uses three specialized Isaacus models: Kanon 2 Enricher for building knowledge graphs and identifying entities (persons, organizations, locations, dates, roles), Kanon 2 Embedder (the world's best legal embedding model) for span-level classification, and Kanon Answer Extractor for entity linking and relationship extraction. All three are encoder-based models, not generative models.
Who is this tool designed for?
This tool is specifically designed for lawyers and legal professionals who need to compare key terms across many documents during tabular-based legal analysis and review workflows. It is also suitable for legal tech developers who want to build or extend tabular review capabilities, and organizations seeking a cost-effective open-source alternative to expensive commercial legal AI tools like Harvey, Legora, and Clio.
How do I get started with Isaacus Tabular Review?
To get started, you need an Isaacus account and valid API key obtained through the API quickstart guide. Set your ISAACUS_API_KEY environment variable (using a .env file), install dependencies (fastapi, uvicorn, isaacus, qdrant-client, dotenv, httpx), clone the Isaacus Cookbook for this guide, and run the server code. Open http://127.0.0.1:8000 in your browser to access the web interface for uploading files and adding columns.
What is the ILDGS schema?
ILDGS stands for Isaacus Legal Document Graph Schema, a knowledge graph schema designed to capture most of the core features present in legal documents. Kanon 2 Enricher automatically organizes identified entities and relationships into this schema, which includes persons (natural and corporate), locations, dates, document segments, cross-references, emails, websites, phone numbers, ID numbers, terms, external documents, quotes, and headings with hierarchical relationships.
What is span-level classification?
Span-level classification is a technique where the tool encodes meaningful spans of text (sentences, clauses, phrases, or multi-sentence sections) using Kanon 2 Embedder and compares them against natural language queries framed as labels. Unlike traditional chunking that returns entire chunks, span-level classification operates at finer granularity, returning only relevant text segments above a similarity threshold (default 0.4 cosine similarity) with character offsets for precise location in the original document.
How does entity linking work?
Entity linking uses Kanon Answer Extractor to perform QA extraction for a user query, returning answer spans from the source document. The system then cross References these spans against the ILDGS document and checks which entities (persons, locations, terms) are mentioned within them using span overlap detection. Any matching entities are linked back to the query by saving their unique IDs, effectively creating new relationships on the fly without having to define them in advance.
What vector database does it use?
The application uses Qdrant, a vector database that supports vector search with custom metadata. The tutorial uses Qdrant in-memory mode (':memory:') for simplicity, which stores embeddings temporarily during the session. Each document gets its own Qdrant collection named 'doc_{file_id}' with vectors sized to match Kanon 2 Embedder's output, using cosine distance for similarity measurement.
Can I commercialize this tabular review application?
Yes, the Isaacus Tabular Review guide is open source and explicitly free to use, adapt, extend, and even commercialize. You can replicate the code, clone the Isaacus Cookbook, modify it for your needs, or build commercial products based on it. The only requirement is having an Isaacus account with valid API key to access the underlying Kanon models via API.