Sage2 by Levanto Labs

Product Introduction

1. Definition: Sage2 by Levanto Labs is an advanced runtime intelligence API designed for developers to extract structured typed fields from natural language and execute real-time streaming decisions. Categorized within the AI developer tools and Natural Language Processing (NLP) runtime model space, Sage2 provides programmatic bricks that translate unstructured human conversation into actionable application state.

2. Core Value Proposition: Sage2 exists to eliminate the friction between unstructured conversational inputs and deterministic software logic. By offering specialized APIs for extraction and streaming decisions, the platform empowers developers to build applications that instantly understand user intent, track changing requirements, validate conditions dynamically, and automate downstream workflows without relying on cumbersome prompt engineering.

Main Features

1. Extraction API: Sage2 parses unstructured natural language inputs and extracts typed fields such as text, numbers, money, dates, times, and durations. Developers define the expected schema with explicit definitions, and the API returns typed values accompanied by source text mapping and probability scores. It supports ISO 8601 standard formats (e.g., converting durations like "over 60 months" into P60M) and handles null selections for missing data points, enabling applications to prompt users for specific missing information before tool execution.

2. Streaming Decisions API: Using WebSocket connections (wss://sage.levanto.ai/v1/stream), Sage2 evaluates conversational context dynamically as text arrives in real time. Developers can configure up to 32 parallel boolean or categorical questions with custom instructions to track shifting user intent—such as detecting explicit offer acceptance versus objections. The streaming engine maintains up to 16,384 tokens of cumulative context, processing incremental updates so that client applications only need to transmit newly generated message segments.

3. Extended Runtime Capabilities: Sage2 integrates advanced cognitive models featuring built-in reasoning modules, web grounding for real-time information retrieval, and beta vision support for multimodal data processing. Furthermore, enterprise deployments can leverage confidential computing executions—such as running Sage on NEAR AI inside trusted execution environments (TEEs)—to securely process sensitive company data and ensure strict privacy compliance.

Problems Solved

1. Pain Point: Traditional Large Language Model (LLM) implementations output unstructured strings that require fragile regex parsing, custom JSON-mode wrappers, or unreliable function-calling loops to drive application logic. This causes systemic application failures when users change their minds, introduce ambiguities, or provide partial data during multi-turn conversations. Sage2 solves this by decoupling extraction and decision logic into dedicated, deterministic runtime APIs that systematically track null values, probabilities, and shifting contextual states.

2. Target Audience: Full-stack software engineers, conversational AI developers, backend architects, and enterprise automation engineers building intelligent agents, customer support workflows, automated document processors, and e-commerce quoting systems.

3. Use Cases: Automating dynamic financing and loan application workflows where users iteratively negotiate terms, budgets, and vehicle types; managing multi-turn customer support escalation paths that require real-time validation of user intent; executing programmatic tool selection based on conditional logic derived from conversational streams; and processing confidential enterprise documents requiring secure, auditable data extraction.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Exceptional precision in extracting typed fields with exact source-text mapping; low-latency streaming decision updates over WebSockets without reprocessing entire conversation histories; robust support for complex data types (money, durations, ISO dates); native handling of null values and confidence probabilities; and enterprise-grade confidential inference options.
   • Cons: Limited to 32 concurrent questions and a 16,384-token context window per streaming session; requires developer implementation of downstream code handling for low-confidence or null extractions; represents a specialized paid SaaS dependency outside of standard self-hosted open-source model weights.

2. Key Alternatives & Differentiation:

   • Alternative 1: OpenAI Function Calling / Structured Outputs: While native LLM provider structured outputs can enforce JSON schemas, they lack real-time conversational streaming decision logic, stateful multi-turn question evaluation, and native duration/money typing tied to exact character positions.
   • Alternative 2: Custom Regex and Rule-Based Parsers: Traditional parsing is brittle and fails against natural language variations. Sage2 provides probabilistic semantic extraction combined with dynamic state tracking that adapts gracefully to changing user inputs.
   • Differentiation: Sage2 uniquely bridges raw natural language and programmatic execution by separating extraction and decision-making into dedicated runtime primitives, offering superior developer ergonomics for building reactive conversational agents.

Frequently Asked Questions (FAQ)

1. What data types does the Sage2 Extraction API support? The Sage2 Extraction API supports typed field extractions including text, numbers, money, dates, times, and ISO 8601 durations (e.g., converting natural language timeframes into machine-readable duration codes like P60M).

2. How does the Streaming Decisions API handle evolving conversation contexts? The Streaming Decisions API connects via a WebSocket protocol to accept incremental text frames. As new conversational data arrives, Sage updates its evaluation against up to 32 pre-configured questions, maintaining up to 16,384 tokens of cumulative context to track whether user intent or agreement status has changed.

3. Can Sage2 handle missing or incomplete user information? Yes. When a user fails to provide a requested detail (such as a loan term), Sage2 returns a null value for that specific field while providing the probability and source text. This allows application code to explicitly identify missing parameters and prompt the user for clarification before triggering backend actions.

4. Is Sage2 suitable for handling sensitive enterprise data? Yes. Sage2 supports advanced enterprise security features, including running inference inside trusted execution environments (TEEs) such as integration with NEAR AI for confidential data classification and automated decision-making on sensitive pipelines.