Synapse · Platform
Let your clients ask the questions themselves.
Put retrieval where your clients already are — a branded widget on their own site, an agent in their Slack, a portal they log into. Each one scoped to their documents and nobody else’s.
In private preview — every plan starts with a conversation.
Ways to hand it over
The same project, reachable however your client prefers to ask.
Website Widget
Embedded on your client’s own site and branded per link. One account can run several — a support bot and an onboarding bot are the same product with different faces.
Slack Agent
Answers in the thread where the question was asked. The conversation lands in the inbox and on the invoice like any other.
Shared Portal
A place to log in and ask, for clients who would rather not embed anything. Share a project by email with a role attached.
Reachable through the API for teams whose clients live there rather than in email.
REST API
Upload, search and chat as three calls. Everything else is a variation on them.
Read-only Links
Hand someone a link to a project without giving them an account or the underlying files.
How It Works
Create a project
One per client, or one per product surface. Retrieval settings are tuned per project.
Upload documents
Parsing, chunking and embedding run behind a queue; status streams back live as each document becomes ready.
Publish it
Drop in the widget, connect the Slack workspace, or share a link. Your client asks; the answers cite their own documents.
Enterprise from the foundations
When a client asks how their documents are kept apart from another client’s, the answer should be a structure rather than a promise.
A Postgres schema per client
Every tenant’s projects, documents, chunks and vectors live in their own schema, created from a template and reached only through a validated identifier. Isolation you can point at, not a policy you have to take on trust.
Roles, checked in one place
Guest through admin, with rows scoped by team underneath. Access is asserted by shared helpers rather than at each handler, so a new endpoint cannot quietly forget.
Hybrid retrieval
Dense vectors and sparse full-text run together and are fused by reciprocal rank, then optionally reranked. Depth, threshold and rerank are set per project, not globally.
Accounted for
Ingestion is a queue with retries and a live status feed. Every completed job records a usage event, so what a client used and what they are billed are the same number.
Observable
Traces, metrics and an evaluation harness, so you can see what retrieval actually returned rather than guessing from the answer.
Self-hosted if you need it
A Helm chart and your own Kubernetes, for firms who need the whole thing inside their own perimeter.
Private preview
Synapse is not self-serve yet. Tell us what you would put in front of your clients and we will get you a key — a checkout that does not work would be worse than an honest queue.
Request access