Product Introduction
Definition: Valori (https://valori.systems) is a deterministic vector infrastructure and vector database with built-in graph support, positioned as an AI memory layer that unifies vector search, knowledge graphs, deterministic state, and cryptographic verification. It exposes an API-first architecture through REST endpoints such as POST /v1/search, plus official Python SDK (valori-client) and TypeScript SDK (@valori/client-sdk). The platform targets AI systems that need predictable, auditable infrastructure for RAG, agents, search, recommendation, and anomaly detection workloads.
Core Value Proposition: Valori exists to solve the "what state produced this result?" problem in vector systems. Instead of backup-and-restore with approximate recovery, Valori makes memory reproducible, replayable, and independently verifiable. Its deterministic kernel uses Q16.16 fixed-point math rather than floating point, so identical inputs produce identical outputs across x86 and ARM machines. Every write commits to a BLAKE3-chained log and returns a receipt containing pre- and post-write state roots and log sequence, enabling auditable vector search and graph relationship queries.
Main Features
Deterministic Fixed-Point Vector Engine: Valori's core computation uses Q16.16 fixed-point arithmetic instead of floating point. This is the technical basis for reproducible vector search across heterogeneous hardware, including x86 and ARM. The vector engine supports collections, records, similarity search, and nearest-neighbor retrieval through a consistent REST API contract. A search request can target a collection with a query vector and k value, returning result IDs and scores. This makes Valori suitable for RAG retrieval, recommendation ranking, semantic search, agent memory, and distance-based anomaly detection. Because the computation is deterministic, the same inputs are designed to produce the same outputs on any machine.
Cryptographically Verifiable State with BLAKE3: Every insert or write in Valori commits to a BLAKE3-chained log. The system returns a receipt with the pre-write and post-write state roots plus the log sequence. This turns state into something that can be independently verified rather than merely backed up. Recovery and replay are designed to restore from snapshot and event log to a byte-identical hash. The trust layer is auditable and deterministic, which matters for AI infrastructure where knowing the exact state behind a search result, recommendation, or agent memory retrieval is critical. The BLAKE3 verification mechanism supports reproducibility claims and audit trails.
Unified Vector and Graph Memory Layer: Valori combines vector search with built-in graph support for relationship queries. Instead of stitching together separate vector databases and graph databases, it provides a unified engine for storing, indexing, searching, and connecting data. Projects act as isolated workspaces for data, models, and applications. Collections organize records; vector search handles similarity; graph capabilities handle relationships. This unified design is intended for AI memory that must be queried by meaning and by connection. It supports RAG, recommendation systems, AI agents, advanced search, data analysis, and anomaly detection.
API-First Access with Official SDKs: Valori provides a REST API, Python SDK, and TypeScript SDK, allowing developers to move from an API key to vector search in minutes. The documented API includes endpoints such as POST /v1/search with collection, query, and k parameters. The response shape is consistent and documented. cURL, Python, and TypeScript quickstarts are available. The open-source Valori Kernel is hosted on GitHub, and the project includes an ArXiv paper (2512.22280), benchmarks, changelog, API reference, and documentation. This reduces infrastructure overhead for teams that want to start building without running a full vector database stack.
Valori × Jev Decision Layer: Valori can reconstruct state that matters, and Jev evaluates it. The Valori × Jev integration is described as moving from memory to decision without an LLM in the middle. Valori produces compiled state involving state, graph, and time; Jev produces decisions with confidence-style outputs such as TRUE 94% and HIGH 91%. This architecture is relevant for systems that need deterministic evaluation of retrieved state rather than probabilistic LLM mediation.
Problems Solved
Pain Point: Traditional vector databases and AI memory systems often make it difficult to know exactly what state produced a given search result. Recovery usually means "probably fine," not provably identical to what existed before. Floating-point non-determinism can cause different results across x86 and ARM machines. Data may be split across separate vector and graph systems, increasing integration complexity. Auditability, reproducibility, and verifiable state are hard to guarantee. Valori addresses these problems with deterministic computation, cryptographic receipts, BLAKE3-chained logs, replayable state, and unified vector-plus-graph queries.
Target Audience: Valori is built for AI engineers, ML platform teams, backend developers, data engineers, search engineers, and recommendation system developers. It is also relevant to Python and TypeScript developers building RAG pipelines, autonomous agents, semantic search, and anomaly detection. Teams in regulated or high-stakes industries such as legal, finance, healthcare, HFT, ecommerce, and stock analysis may need auditable, deterministic vector infrastructure. Compliance, security, and infrastructure teams that require verifiable state and reproducible AI behavior are also target users.
Use Cases: RAG: Store document chunks as vectors and retrieve nearest matches to ground generation with relevant, current context. Recommendation systems: Search a collection for nearest neighbours to a user or item vector to rank by similarity at speed. AI agents: Insert and search vectors as long-term, queryable memory across sessions. Advanced search: Query by meaning, not just keyword match, and optionally blend term-frequency reranking. Data analysis and anomaly detection: Find records far from their neighbours in vector space. Verifiable AI infrastructure: Maintain BLAKE3 receipts, state roots, and replayable logs for audit and reproducibility. Graph-enhanced memory: Use relationship queries alongside vector similarity for connected data.
Unique Advantages
Differentiation: Valori differentiates from traditional vector databases by treating determinism and verification as core infrastructure rather than afterthoughts. Many vector systems rely on floating-point computation and backup-based recovery. Valori uses Q16.16 fixed-point math and BLAKE3-chained state logs with pre- and post-write state roots. It also unifies vector search and graph relationship queries in one memory layer, reducing the need to operate separate systems. The open-source kernel, ArXiv paper, documented REST API, Python SDK, and TypeScript SDK provide transparency and developer accessibility. Claims are backed by benchmarks and a changelog rather than only marketing language.
Key Innovation: The key innovation is deterministic state for vector infrastructure. By computing in Q16.16 fixed-point instead of floating point, Valori aims to produce identical outputs on any machine, x86 or ARM. Every write commits to a BLAKE3-chained log and returns a receipt with state roots and log sequence. State can restore from snapshot and event log to a byte-identical hash. This makes AI memory reproducible and independently verifiable. The Valori × Jev integration extends the concept from memory reconstruction to deterministic decision evaluation without an LLM in the middle.
Frequently Asked Questions (FAQ)
What is Valori? Valori is a deterministic vector infrastructure and vector database with built-in graph support. It combines vector search, knowledge graphs, deterministic state, and verification into one memory layer for AI. Developers use it to build RAG, agents, search, and recommendation systems with memory that can be reproduced and independently verified. It offers a REST API, Python SDK, and TypeScript SDK.
How does Valori make vector search deterministic? Valori's kernel computes in Q16.16 fixed-point arithmetic instead of floating point. This means the same inputs produce the same outputs on any machine, whether x86 or ARM. Every write also commits to a BLAKE3-chained log and returns a receipt with pre- and post-write state roots and log sequence. State restores from snapshot and event log to a byte-identical hash, enabling replay and verification.
Is Valori open source and does it have SDKs? Yes. The Valori Kernel is open source on GitHub at varshith-Git/Valori-Kernel. Valori provides a REST API, an official Python SDK (valori-client on PyPI), and an official TypeScript SDK (@valori/client-sdk on npm). Documentation, API reference, benchmarks, changelog, and an ArXiv paper (2512.22280) are available.
What use cases does Valori support? Valori supports RAG document retrieval, recommendation systems, AI agent long-term memory, advanced semantic search, data analysis and anomaly detection, and graph relationship queries. It is designed for workloads where knowing exactly what state produced a result matters, including legal, finance, healthcare, HFT, stocks, and ecommerce.
How does Valori compare to traditional vector databases? Traditional vector databases often rely on floating-point computation and backup-based recovery, which can make exact state reproduction difficult. Valori uses deterministic Q16.16 fixed-point computation, BLAKE3-chained state logs, and cryptographic receipts. It also adds built-in graph support and a verification trust layer, so vector search and relationship queries can operate over auditable, replayable state.
