Hybrid retrieval
Lexical and vector search with a reranker tuned on your own labelled pairs, not a generic benchmark.
Features
Retrieval engineered like infrastructure
The parts that separate a demo from something legal, clinical and editorial teams will actually use.
Citations first
Every claim links to a passage. Answers without adequate support are withheld, not softened.
Eval harness
Your question set becomes a regression suite. No prompt or model change ships without a green run.
Permission-aware
Document ACLs are enforced at retrieval time — nobody sees a passage they could not open at source.
Model choice
Anthropic, OpenAI, or an open model in your VPC. Swap without rewriting the retrieval layer.
Usage analytics
See what people ask, where retrieval fails and which documents are quietly load-bearing.
In practice
Where Signal earns its keep
Pick a workflow to see how it is configured.
Pricing
Priced per team, not per seat scramble
Every plan includes onboarding, migration support and a named engineer. 30-day trial, no card.
Workspace
One team, one corpus, cited answers from day one.
$890/mo
- Up to 250k documents
- 3 source connectors
- Cited answers + abstention
- Email support, next business day
Most chosen
Business
Multiple corpora with permissions and evals in CI.
$2,600/mo
- Up to 5M documents
- All connectors + ingest API
- Permission-aware retrieval
- Eval harness in your CI
- SSO, SCIM, audit log
Enterprise
Your infrastructure, your models, our engineers.
Custom
- On-prem or VPC deployment
- Bring your own model
- Custom evaluation programme
- Named engineering pod
- Security review and DPA
Questions
Before you commit
How is this different from a RAG demo?
Demos optimise for a good first answer. Signal optimises for the hundredth — abstention, permission enforcement, citation precision and a regression suite that runs before every release.
Do you train on our documents?
Never. Your corpus is isolated, embeddings stay in your tenant, and no data is used for training by us or the model providers we call.
Which models can we use?
Anthropic and OpenAI hosted, or an open-weights model in your own VPC. Retrieval is model-agnostic, so switching is a config change.
How long is implementation?
Two to four weeks: connector setup, corpus ingest, then a calibration round where we build the eval set with your experts.
Signal · AI
Bring your hardest questions
We will index a sample of your corpus and run your own questions against it before you sign anything.