RAG Infrastructure

Build RAG That
Stays Good

Experiment with strategies, version your corpus, and catch quality regressions before your users do. Start free with 15 documents in under a minute.

Upload docs · run strategies · compare results in minutes

Integrates with:
PineconeQdrantAWSMilvus
Step 01 — Upload
Version historyprod-docs
v1

Initial corpus

1,204 docs

Mar 1

v2

Added Q1 content

1,891 docs

Mar 14

v3

Pruned stale pages

1,743 docs

Mar 22

v4

Latest sync

1,910 docs

Mar 30

current

SDK

client.datasets.upload(
  name="prod-docs",
  files=["./docs/**/*.md"]
)
# → version created automatically
01Upload

Connect your corpus.
Versions are automatic.

Every sync or upload creates a new version automatically. No setup, no opt-in. Your full corpus history goes back to the first upload, so you always have a baseline to compare against.

Every upload creates an immutable version snapshot
Full history back to v1, retrievable at any time
Compare any two versions to see what changed
Roll back to any previous corpus state instantly
Step 02 — Quality
02Quality

Catch corpus drift
before your users do

Every new version automatically triggers centroid shift, norm shift, and distribution checks. You get a clear pass or fail before anything reaches production. No instrumentation to write.

Runs automatically on every version create
Centroid shift catches directional corpus drift
JS divergence catches structural distribution changes
Pass or fail with raw metric values, not just status
Quality — v4 vs v3
All checks passed

Centroid Shift

Semantic center is stable

0.08

threshold 0.10

Norm Shift

Vector magnitudes unchanged

0.06

threshold 0.15

Vector Validity

No NaN or zero vectors

100%

threshold 100%

Duplicate Rate

Near-zero redundancy

1.2%

threshold 5%

What a failure looks like

!

Centroid Shift exceeded threshold

Large semantic shift detected — review new batch

0.31

threshold 0.10

Step 03 — Deploy
Deploy history — prod-docs
Open dashboard
v1

ada-002 · 256t chunks

Mar 1

Recall 0.71

rolled back

v2

3-small · semantic chunks

Mar 14

Recall 0.84

superseded

v3

3-large · hybrid search

Mar 22

Recall 0.91

superseded

v4

3-large · late chunks · hybrid

Mar 30

Recall 0.94

live

Any version can be promoted back to live. Re-embedding not required.

03Deploy

Promote, rollback,
or branch. Your history is always there.

Every retrieval config that has ever shipped is stored. Compare recall across versions, promote the winner, or rollback to any prior state in seconds. No re-embedding required.

Branch a corpus version to test a new config safely
Compare Recall@K across every version side by side
Promote any version to production with one call
Rollback in seconds — no re-embedding, no data loss
Get Started
For Teams Running Retrieval in Production

Your pipeline is one setup away
from being observable

Start with the free RAG Lab. No signup needed. Upload 15 documents, generate a gold set, and compare strategies in minutes.

ExperimentUploadQuality checksDeploy