Hybrid sparse-dense retrieval, now generally available

The reference layer your RAG pipeline needs

Normalize, chunk, embed, and retrieve across heterogeneous corpora. One API, compatible with LangChain, LlamaIndex, and any custom stack.

Get API key Read the quickstart

ragref retrieve

$ curl ragref.com/v1/retrieve \ -d '{"corpus":"support-docs", "query":"reset 2FA device","rerank":true}' # 200 OK · 41ms { "matches": [ { "score": 0.914, "doc": "security/2fa.md" }, { "score": 0.877, "doc": "account/devices.md" } ], "latency_ms": 41 }

10Btokens per corpus

100kqueries per second

<50msp99 retrieval latency

99.95%indexing uptime

Built for production RAG

Retrieval quality and index durability, without operating your own vector infrastructure.

Retrieval

Hybrid search that holds up out of domain

BM25 sparse and dense vectors are fused with reciprocal rank fusion, then optionally passed through a cross-encoder reranker before anything reaches your context window.

Sparse + dense fusion with tunable weighting
Cross-encoder reranking on retrieved candidates
Citation provenance on every match

dense0.72

sparse0.58

fused0.91

reranked0.96

Chunking

Semantic chunking that preserves context

Paragraph, sentence, and token-boundary strategies with semantic overlap detection, so a split never severs a reference from the claim it supports.

Configurable boundaries and overlap
Structure-aware splitting for code and tables
Deterministic, replayable ingestion

Two-factor authentication adds a second step at sign-in. To reset a lost device, an account owner opens Settings, selects Security, and confirms identity by email before a new authenticator can be enrolled.

overlappreserved

A small, predictable API

Everything is scoped to a corpus. Everything returns provenance.

Corpora

POST/v1/corporaCreate a new document corpus

POST/v1/corpora/{id}/ingestIngest and chunk documents

GET/v1/corpora/{id}/statsCorpus metadata and index health

Retrieval

GET/v1/retrieveHybrid sparse-dense retrieval

POST/v1/rerankCross-encoder reranking

GET/v1/embedEmbeddings, OpenAI-compatible

Live in five lines

Install the client, create a corpus, ingest a directory, and retrieve. No infrastructure to stand up.

# pip install ragref
from ragref import RAGRefClient

client = RAGRefClient(api_key="rr_live_...")

corpus = client.corpora.create(name="support-docs", chunking="semantic")
corpus.ingest(path="./docs/", recursive=True)

results = client.retrieve(
    corpus_id=corpus.id,
    query="How do I reset my 2FA device?",
    top_k=8, rerank=True,
)
      

Pay for what you retrieve

Indexing is free at every tier. Billing tracks retrievals and stored tokens, nothing else.

PRO · MOST POPULAR

$149/mo

100M tokens indexed

500,000 retrievals/mo
Unlimited corpora
Hybrid search + reranking
Webhook events
Priority email support
Retrieval analytics

Get API key

STARTER

$0/mo

Up to 1M tokens indexed

10,000 retrievals/mo
2 corpora
Basic chunking
Community support

Start free

ENTERPRISE

Custom

Unlimited scale

Unlimited tokens & retrievals
Private VPC deployment
Custom embeddings
SLA + dedicated CSM

Talk to us