Application Foundation · HRAG

One query. The right retrieval method.

A multi-tenant RAG service that unifies vector, BM25, hybrid, and GraphRAG retrieval behind one API. A routing layer decides which method to use by reading the query itself.

Status
Prototype
Last verified
2026-07-18
Access
TODO_CONTENT_CONFIRMATION

The problem it solves

A single retrieval method (pure vector search, for example) doesn't fit every query type: exact-term lookups, relationship questions, and broad summarization each call for a different retrieval strategy.

HRAG aims to remove that repeated decision from the application by offering a routing layer that picks the right strategy for the query type.

How it works

  1. Send your documents to your workspace collection — synchronously or asynchronously.

  2. When a query arrives, the routing layer reads the query text.

  3. The appropriate lane — vector, BM25, hybrid, or graph — runs.

  4. The answer returns along with the method that was used.

Verified capabilities

Routing by query type

Exact-term queries are routed to BM25, relationship questions to the entity-centric graph, and broad summary requests to the community-centric graph.

A built-in measurement layer

LLM-free metrics such as hit-rate, MRR, and nDCG are designed to run in CI as a regression gate; a nightly RAGAS LLM-judge run additionally tracks the quality trend.

Tenant-level isolation

Every tenant runs in its own workspace: row-level security in Postgres, a single mandatory access gateway in Neo4j, and separate prefixes in Redis and MinIO aim to keep tenants apart.

Two layers of duplicate protection

Two independent dedup layers exist: an Idempotency-Key at the job level, and a content hash at the document level.

Architecture and integration

HRAG serves vector + graph (Neo4j) + BM25 hybrid retrieval, routing, and reranking behind a single multi-tenant M2M API.

Asynchronous ingest delivers at-least-once over a message queue; sibling projects in the ecosystem connect to HRAG with a single API key.

Trust and control model

Tenant isolation is designed with row-level security, a single access gateway, and separate prefix strategies — an attempt to reach another tenant's data is meant to be handled without even revealing that the resource exists.

Automated measurement (the hit-rate/MRR/nDCG regression gate and RAGAS runs) is part of the development process; this page doesn't claim a fixed, up-to-date figure for the full test set's status against real infrastructure — coverage keeps expanding.

Known limits / areas for improvement

Access model not yet settled

The access model hasn't been confirmed by the product owner yet.

No up-to-date test report published

Because test count and coverage change over time, this page doesn't state a fixed number; an up-to-date, independently verified report hasn't been published yet.

Prototype maturity level

The service is at the prototype stage; multi-tenant scale testing is ongoing.

Evidence

The CI regression gate and nightly RAGAS runs are part of the internal development process; a publicly available, up-to-date test report or evidence link hasn't been published yet. This section will be updated once one is published.

Don't build your RAG. Connect to the right method.

Request a technical call