Key facts
| RAG endpoints | POST /v1/embeddings, POST /v1/rag/collections and POST /v1/rag/query |
| Embedding models | plugsky-embed-v1 (1536 dimensions, OpenAI-ada compatible) and plugsky-embed-large (3072 dimensions) |
| Retrieval modes | Keyword, vector and hybrid search with optional cross-encoder reranking |
| Citations | Every query returns ranked chunks with source attribution |
| Ingestion | Documents are chunked, embedded and indexed automatically per collection |
| Batch limits | Up to 2,048 inputs per embeddings request, max 8,191 tokens each |
| Data handling | Per-collection encryption at rest; API data is not used to train models |
| Deployment | Hosted, VPC, on-prem and air-gapped options; 30+ models behind one API |
TL;DR
- Treat each corpus as a product with an owner, freshness target and access rule.
- One OpenAI-compatible plane covers ingestion, retrieval and generation.
- Route cheap classification to small models and pin stronger models for sensitive answers.
- Choose the data plane per corpus: region-locked, VPC, on-prem or air-gapped.
- Keep the embedding interface stable so vector stores remain swappable.
How it works, step by step
- Inventory corpora, owners, refresh cadence and sensitivity classification.
- Map each corpus to an allowed deployment plane and a collection boundary.
- Decide retrieval mode per corpus: keyword, vector or hybrid with reranking.
- Define model routing rules: default, escalation and pinned versions for regulated answers.
- Specify key custody, audit fields, retention and SIEM export with security.
- Prove retrieval quality on one corpus with a shared evaluation set.
- Standardise an ingestion contract and open the service to further teams.
Original data
Try it yourself
Open the RAG architecture builder →
Why enterprise architects need RAG grounded in their own content
Architects are asked to place RAG on a diagram and defend it. Four questions decide the shape: where each corpus lives and who owns it, whether retrieval quality needs vector or hybrid search, which data plane is allowed for each corpus, and which model is permitted to answer. Get those wrong and every later change is a migration.
Treat corpora as products with owners, refresh cadences and access rules. A collection per bounded domain — HR policy, product documentation, regulatory notices — keeps retrieval precise, permissions clean and evaluation practical.
The Plugsky RAG stack: embeddings, collections, queries
The retrieval pipeline is three calls: embed, collect, query. POST /v1/embeddings converts text to vectors using plugsky-embed-v1 (1536 dimensions, OpenAI-ada compatible) or plugsky-embed-large (3072 dimensions); POST /v1/rag/collections stores and indexes documents with per-collection encryption at rest; POST /v1/rag/query returns ranked chunks with source attribution. Keyword, vector and hybrid retrieval are all available, with optional cross-encoder reranking when precision matters more than raw recall. For architects the decision is which layer to own: managed collections shorten delivery, while standalone embeddings let you keep an existing vector store without changing the model interface.
The controls enterprise architects should configure first
A defensible reference topology puts ingestion, retrieval and generation in one OpenAI-compatible plane: documents are chunked and embedded into collections, queries return ranked chunks with citations, and any chat model composes the answer. Model routing across 30+ models lets you send cheap classification to a small model and sensitive answers to a pinned, stronger one.
Residency, keys and audit complete the picture: region-locked planes, VPC, on-prem or air-gapped deployment; BYOK via AWS KMS, Azure Key Vault, HashiCorp Vault or an on-prem HSM; per-request logs with SIEM export and retention up to 7 years. Review model versions at the same cadence as your other dependencies.
A practical pilot path
Phase the rollout: prove retrieval on one corpus, then standardise the ingestion contract, then open the service to more teams through scoped keys and quotas. Keep the embeddings interface stable even if you change vector stores later, and hold the evaluation set as a shared asset so changes are measured, not argued. Prototype on the free plan and use the 14-day full-access trial for a wider evaluation. See the live pricing page for current plans.
Honest comparison
| Capability | Plugsky | Assembling your own stack | Keyword search only |
|---|---|---|---|
| Retrieval | Keyword, vector and hybrid with optional reranking | You integrate and tune each component | Keyword matching, no semantics |
| Citations | Ranked chunks with source attribution | You build attribution yourself | Document links, not passages |
| Embeddings | plugsky-embed-v1 (1536d) and plugsky-embed-large (3072d) | Model choice spread across vendors | Not applicable |
| Deployment | Hosted, VPC, on-prem and air-gapped | Depends on every component | Usually on-prem already |
| Data handling | Per-collection encryption; API data not used to train models | You audit every vendor | Existing controls apply |
| Time to first answer | One collection and a query call | Weeks of integration work | Days, with lower answer quality |
Frequently asked questions
How many collections should we create?
One per bounded domain with a clear owner. Separate collections keep retrieval precise, make per-collection encryption and access rules meaningful, and let teams evaluate independently.
How does model routing fit a RAG design?
Retrieval returns chunks independent of the model; routing decides which model composes the answer, so you can send routine work to efficient models and pin stronger models where answers need review.
Do we have to change our application code?
No. The RAG endpoints are OpenAI-compatible, so existing SDKs, prompts and evaluation harnesses keep working when you point the base URL at Plugsky.
Is there a free plan?
Yes — plugsky-micro and plugsky-lite with 2 API keys and no credit card. A 14-day full-access trial is also available.
How is pricing structured?
Self-serve plans are flat monthly with unlimited fair-use usage and no per-token charges or overage fees. See the live pricing page for current plans.
Does Plugsky train on our data?
Plugsky states that API data is not used to train models, and collections are encrypted at rest; on Enterprise you can keep everything inside your own VPC.