Voxell Answers

Production RAG in one call.

A question goes in. A grounded, cited answer comes out. The retrieval pipeline between them (chunking, embedding, indexing, reranking, synthesis) is Voxell's problem now, not yours.

already embedding? → your corpora

want this without an API? → Spaces publishes it as a page instead

answers · first call● live
$ curl https://api.voxell.ai/v1/answers/runbooks/query \ -H "Authorization: Bearer $VOXELL_KEY" \ -d '{ "query": "How do we rotate the signing keys?", "mode": "answer" }'
// full response: paste-and-go, nothing trimmed { "mode": "answer", "query": "How do we rotate the signing keys?", "answer": "Rotate from the admin console under Keys → Rotate. The previous key stays valid for 24 hours while scheduled jobs re-authenticate.", "chunks": [ { "vector_id": "vec_8f2a1c", "doc_id": "runbooks/key-rotation.md", "chunk_index": 7, "score": 0.91, "text": "To rotate a signing key: open Admin → Keys → Rotate. The previous key stays valid for 24h while…" } ], "retrieval": { "health": "healthy", "top_score": 0.91 } }
#1 on MTEB English v2 · Ingot-8B-R3 · 75.98 Mean(Task) · 41 tasks
87ms retrieval, against a 300ms market average
Free at Turbo
The pipeline tax

Agents need answers. RAG stacks return chunks.

Every agentic pipeline that touches a vector database inherits the same build: a chunking strategy, an embedding bake-off, index provisioning, hybrid-search weights, a reranker, context assembly, citation plumbing, and an eval harness to prove any of it works. Each stage is latency, cost, and failure surface inside your agent's loop, and none of it is your product.

Answers is built for the pipeline that calls it: deterministic JSON, citations your agent can verify before it acts, and a grade on every source, so confidence is a number, not a vibe. Ask with mode=vectors when the agent wants raw chunks; mode=answer when it wants the finished object.

chunking_strategy.py
embedding_model_bakeoff.md
vector_db_provisioning.tf
hybrid_search_weights.yaml
reranker_config.json
context_assembly.py
citation_extraction.py
retrieval_eval_harness/
curl https://api.voxell.ai/v1/answers/runbooks/query

The retrieval stack, replaced by one endpoint your agent calls as a tool.

Under the call

Forge vectors in. Grounded answers out.

Embed

The #1 model does the reading

Forge encodes your corpus and every query with Ingot-8B-R3, first on MTEB English v2. Retrieval quality is the ceiling on answer quality. We set the ceiling.

Retrieve

Edge speed or Max precision

Standard corpora are edge-served for latency. Max corpora run precision retrieval on Voxell GPUs. Both return the passages that actually answer, ranked.

Answer

The finished object

Retrieved passages become one grounded response with citations attached, ready for an agent to verify and act on, not a context window to babysit.

Sovereign inference

rolling out

Embedding and retrieval already run end to end on Voxell-operated infrastructure. Generation joins them next: a sovereign model on Voxell GPUs, so a query and its corpus never transit a third-party model API. One trust boundary, from document to answer.

Your corpus

Bring the corpus you already have.

Migration starts with an assessment, not a rewrite. Point Answers at your existing documents; Forge ingests them and VQS grades every passage for how well it can support an answer.

You see which sources are weak before your agent finds out in production. Then fix, replace, or drop them with evidence instead of guesswork. Proof, not vibes.

Assess your corpus →

runbooks/key-rotation.md 0.91
runbooks/incident-response.md 0.87
wiki-export/legacy-faq.html 0.42

VQS per passage, per corpus. The weak third of most corpora is where wrong answers come from.

Pricing

Pricing that doesn’t tax idle data.

You pay to embed a document once, then a flat toll per query. Not monthly rent on every vector sitting in storage, so a corpus you rarely ask about costs about nothing to keep.

That means the bill tracks how much you use the corpus, not how big it got. See what’s included →

New to RAG?

Why not just paste it into a chat?

For one document and one question, a chat window is fine. It falls apart the moment you have more documents than fit in a prompt: a support team with 50,000 articles, a contract set spanning years, a product answering from a live knowledge base. You cannot paste all of it into a chat.

Answers pre-processes your whole corpus into searchable meaning once, then retrieves only the few passages that matter when a question arrives. The model sees the right context instead of everything, so it is faster, cheaper, and it scales. And because retrieval quality is the ceiling on answer quality, it runs on embeddings that are benchmarked rather than assumed.

First answer in five minutes.

Create a key, post a document, ask a question. The first call is the demo.

Get an API key

$ curl https://api.voxell.ai/v1/answers/{corpus}/query · free at Turbo