Papr Graph

Product Introduction

1. Definition: Papr Graph is a graph-native embedding API and retrieval enhancement layer. Technically, it is a middleware service that sits between an embedding model and a vector database to transform standard semantic embeddings into multi-dimensional, graph-aware representations.
2. Core Value Proposition: Papr Graph exists to solve the critical problem of semantic retrieval inaccuracy in RAG (Retrieval-Augmented Generation) and AI agent systems. Its core value is enabling vector search to retrieve answers based on correctness, relevance, and contextual intent—not just cosine similarity—by encoding relational, temporal, and topical signals directly into the embedding space.

Main Features

1. Graph-Aware Embedding Transformation: The papr.transform endpoint takes a document and its standard vector embedding (from any provider like OpenAI, Cohere, or Voyage AI) and rotates it within the vector space. This rotation is based on up to 14 pre-defined or custom relational dimensions, such as document recency, approval status, entity scope (e.g., subsidiary, region), and topical lineage.
2. Intent-Based Reranking: The papr.rerank endpoint operates at query time. It takes the top-K candidate documents retrieved via standard cosine similarity from a vector DB and re-scores them by matching the query's inferred intent against the graph signals (phases) attached to each candidate. This ensures the most contextually correct result ranks highest.
3. Stateless, Secure API Architecture: Papr Graph processes data in-memory with zero retention; no embeddings, documents, or queries are stored or logged. Data is encrypted in transit with TLS 1.3. This architecture ensures data privacy and eliminates vendor data lock-in, as all processed data remains in the user's own vector database.

Problems Solved

1. Pain Point: Traditional vector similarity search often retrieves outdated, incorrect, or contextually inappropriate documents because cosine distance cannot natively encode relational metadata like version history, access permissions, or temporal decay. This leads to AI hallucinations and unreliable RAG pipelines.
2. Target Audience: AI/ML Engineers building production RAG systems, DevOps and Platform teams managing AI infrastructure, and Enterprise Architects in knowledge-intensive industries like finance, healthcare, and legal tech who need precise, audit-aware document retrieval.
3. Use Cases: Ensuring an AI agent retrieves the current, in-force company policy instead of an archived draft; preventing a customer support bot from surfacing draft or region-restricted documentation; improving accuracy in academic or research retrieval where citation recency and authority are critical.

Unique Advantages

1. Differentiation: Unlike standalone vector databases (Pinecone, Weaviate) or embedding models, Papr Graph is a dedicated signal-injection and reranking layer. It enhances, rather than replaces, existing stacks. Unlike simple metadata filtering, it uses graph signals to mathematically rotate the entire embedding space for more nuanced retrieval.
2. Key Innovation: The product's core innovation is the method of "rotating" pre-existing embeddings based on graph-derived signals. This allows teams to gain graph-native retrieval benefits without retraining models, switching vector databases, or manually crafting complex hybrid search rules, achieving state-of-the-art benchmark results as validated by Stanford's STARK evaluation.

Frequently Asked Questions (FAQ)

1. What is Papr Graph and how does it work with my existing vector database? Papr Graph is an API that adds a graph-aware processing layer to your retrieval pipeline. You keep your chosen embedding model (e.g., OpenAI's text-embedding-3) and vector database (e.g., Pinecone, pgvector). At index time, you pass generated embeddings to Papr's transform endpoint to enrich them. At query time, you pass top search results to the rerank endpoint for intent-based reordering.
2. How does Papr Graph improve RAG accuracy compared to simple semantic search? Standard semantic search relies solely on cosine similarity, which can rank outdated or irrelevant documents highly if they are linguistically similar. Papr Graph encodes dimensions like time, approval status, and entity relationships into the vector math itself, enabling the system to prioritize the contextually correct answer over the merely semantically close one, drastically reducing retrieval errors for AI agents.
3. Is my data secure when using the Papr Graph API? Yes, Papr Graph employs a stateless, zero-retention architecture. Your document text and embeddings are processed in-memory for the duration of the API call and are immediately discarded. No data is persisted to logs or databases. All communications are encrypted with TLS 1.3, aligning with enterprise security requirements.
4. What are "graph signals" or "phases" in Papr Graph? Graph signals (called "phases") are the discrete relational dimensions Papr attaches to each embedding. These can be built-in signals like recency, approval, and scope, or custom signals you define. They represent the non-textual, contextual metadata about a document that influences its relevance, which Papr uses to rotate the vector and guide the reranking process.
5. What does the Papr Graph pricing model include? Papr Graph uses a consumption-based model centered on "operations," where one operation is one call to either the transform or rerank endpoint. The Developer tier is free for up to 1,000 operations monthly. Paid plans (Starter, Growth) offer higher operation volumes, custom schemas, priority support, and advanced analytics, with Enterprise plans available for VPC deployment and custom agreements.