> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dataframer.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Release Notes
## **New Features**
### **π‘οΈ Contradiction Check**
A new LLM-based quality gate that catches logically contradictory attribute combinations in sampled data (e.g. "sunny weather" + "heavy rain") and automatically re-samples.
***
### **π Distribution Breakdown**
Distribution analysis now shows three separate views for each property axis:
* **Requested** β what the spec asks for (target percentages)
* **Expected** β what's achievable given the sample count (accounts for conditional probability cascades)
* **Evaluated** β what an LLM classifier determined each sample actually is
The "Expected" calculation now properly accounts for conditional probability cascades instead of naively using base distributions.
***
### **β‘ Progressive Results**
The run detail page now shows samples and cost as they complete in real time, instead of waiting for the entire run to finish.
***
### **π Anonymization API**
Released the first mature version of the anonymization (PII/PHI redaction) API, with full documentation, SDK support, and MCP integration.
***
### **π Faster App Experience**
Faster page loads across the app β initial navigation, run details, and profile pages all render noticeably quicker.
***
### **π Auth Reliability**
Fixed several edge cases that could cause login failures or unexpected logouts.
***
## **New Features**
### **π New PDF Generation Engine**
The PDF pipeline has been rebuilt from the ground up. DataFramer now generates a **unique visual style for each document**βlayout, fonts, colors, structure.
**Highlights:**
* **Automatic visual QA**: Every generated PDF goes through automated revision cycles and quality checks to ensure its visual excellence.
***
### **π’ Calculator Tool**
A sandboxed Python execution environment is now available to the LLM during data generation, ensuring numerical accuracy across all generated tables and documents.
**Highlights:**
* **Arithmetically correct output**: The LLM computes totals, percentages, and cross-checks figures in the document, eliminating hallucinated numbers that don't add up. This happens multiple times both before and after writing them into the document, ensuring extremely low error rates.
***
### **π° New Cost Controls**
New run parameters give you fine-grained control over generation cost and speed, enabling cost savings.
**Features:**
* **One-shot generation**: Generates the entire document in a single LLM call instead of the default outline β sections β concatenation process. Enabled by default β disable it when using weaker models or if the document is too long for even state-of-the-art models.
* **Selective revision types**: Instead of all-or-nothing document revisions, you can now enable only the specific revision passes you need for your use case.
***
### **π― Conformance Filtering**
A new post-generation quality gate that ensures every sample in your dataset actually matches its target specification and desired properties.
**Highlights:**
* **Automatic regeneration**: Documents that clearly violate their target properties are automatically discarded and regenerated from scratch β no manual review needed.
***
### **π PII/PHI Anonymization**
A new experimental tool for redacting sensitive information from existing datasets at scale and at extremely low cost (\~\$0.1 / million tokens).
**Capabilities:**
* **Quality evaluation**: An optional LLM-based evaluation step measures precision, recall, and F1 of the redaction.
* **Broad file support**: Works with CSV, JSON, JSONL, Markdown, and plain text datasets.
***
## **New Features**
### **π PDF Support**
DataFramer now supports PDFs as a first-class file type across datasets and generation workflows.
**Highlights:**
* **Dataset ingestion**: `.pdf` is now accepted anywhere you upload dataset files, alongside `.txt`, `.md`, `.json`, `.csv`, and `.jsonl`.
* **Template prompts for PDFs**: you can pass a prompt to control the visual style of generated PDFs (e.g., "Professional corporate style with blue headers").
***
## **New Features**
### **π New Blog Posts & Studies**
New research and tutorials on the [DataFramer blog](https://dataframer.ai/blog):
* **[How to Generate 50K-Token Documents: Same LLM, Different Results](https://dataframer.ai/posts/long-text-generation-dataframer-vs-baseline)** β benchmark study comparing DataFramer vs. raw Claude Sonnet 4.5 for long-form text generation, with a companion dataset on HuggingFace
* **[Generation of Synthetic Text2SQL Data with 100% Validity](https://dataframer.ai/posts/amplifying-claude-haiku-text-to-sql)** β tutorial on generating diverse verified text-to-SQL samples using DataFramer
***
### **𧱠Databricks Integration**
Full integration with Databricks for data ingestion, generation, and model hosting. See the [Databricks integration guide](/integrations/databricks) for a full walkthrough.
**Capabilities:**
* **[pydataframer-databricks](https://pypi.org/project/pydataframer-databricks)** β new Python package for working with DataFramer directly from Databricks notebooks. Includes `DatabricksConnector` for fetching sample data from Unity Catalog tables and loading generated data back into Delta tables via service principal M2M OAuth.
* **Databricks native models** β Databricks-hosted models can be used for specs, generation, evaluation, and chat. A DataFramer admin configures service principal credentials once in the DataFramer UI, and any team member can then select `databricks/` models without passing credentials in API calls
***
### **π° Cost & Time Estimates**
See estimated cost and generation time before starting a run. Because DataFramer uses an agentic generation workflow with multiple LLM calls per sample, costs were previously difficult to predict. The estimator uses a simulated model of the full generation pipeline to produce forecasts before you commit to a run.
**How it works:**
* Estimates update live as you adjust sample count, model, dataset type, and other parameters on the Create Run page
* Accounts for all stages of generation: outline, content, revision cycles, and evaluation
***
### **π Public API**
Stable public REST API for programmatic access to the full DataFramer workflow β datasets, specs, generation, evaluation, and red-teaming. The API went through a major overhaul to reach a stable, consistent interface.
**Highlights:**
* [Python SDK](https://pypi.org/project/pydataframer/) (pydataframer) with typed methods for every endpoint
* Thoroughly documented in the [API Reference](/api-reference) with Python code examples for every endpoint
***
### **π€ MCP Server**
DataFramer is now available as an [MCP (Model Context Protocol)](https://modelcontextprotocol.io/) server, allowing AI assistants like Claude Code, Cursor, and other MCP-compatible clients to interact with the platform directly.
**Capabilities:**
* Upload datasets, create specs, generate data, and download results β all through natural conversation with an AI assistant
* Unlike the raw API, MCP also provides your AI assistant with detailed instructions on how to use DataFramer effectively β so it can guide you through the entire workflow conversationally
* See [API & MCP](/api-and-mcp) for setup instructions
***
### **π PDF Generation**
Generate synthetic PDF documents with custom styling. Describe the visual style you want (e.g., "professional corporate style with blue headers") and DataFramer generates styled PDFs automatically.
**Capabilities:**
* Full PDF input/output β use PDF seed examples and generate new PDF documents
* Custom styling via a natural language prompt that controls headers, fonts, colors, and layout
***
## **New Features**
### **π± Seedless Generation**
Generate high-quality synthetic data without requiring any seed examples. Simply describe what you want and let DataFramer create it from scratch.
**How to create a spec (blueprint for the data) without uploading examples:**
1. Select "Seedless" as the specification type in the spec creation wizard
2. Provide a spec name and generation objectives
3. Set your target token range (e.g., 2,000-5,000 tokens)
***
### **π₯ Admin Tools**
New internal administration capabilities for managing teams and users.
**Features:**
* Role-based access control with Admin and User roles
* Admins can promote/demote users between Admin and User roles
* Company-wide user visibility and management from the Profile page
***
### **π³ Billing System**
Usage-based billing with transparent pricing and detailed invoicing.
**How it works:**
* Calendar month billing cycles (1st to last day of each month)
* Run Details page now shows the cost of your run
* Failed task cost exclusion - you're not charged for failed runs
* Team and Enterprise plan types
***
### **ποΈ ToolBox - SQL Execution Environment**
Multi-database SQL validation engine for generating high-quality Text-to-SQL datasets.
**Capabilities:**
* Validates both schema DDL and query SQL
* Parallel testing against 3 databases: PostgreSQL, MySQL, SQLite
* REST API integration for programmatic access
***
### **π€ Gemini 3 Pro Support**
Full integration of Google's latest Gemini 3 Pro models across the platform.
**Capabilities:**
* Minimal reasoning mode (gemini/gemini-3-pro-preview) and high reasoning mode (gemini/gemini-3-pro-preview-thinking)
* 1 million token context window
* Available for spec analysis, generation, evaluation, red-teaming, and chat
***