Overview
promptev hosts the Context Engine. You never install it, run it or configure it in code. The SDKs, the HTTP API and the MCP server all read and write the same knowledge base, set up the way this page describes.
| Setting | Where | What it decides |
|---|
| Where the knowledge lives | Workspace: Settings → Context engine | Managed by promptev, or your own database. |
| Your own models | Workspace: Settings → Context engine | Embedding, vision (OCR), extraction and rerank models, with your own keys. |
| Code execution | Workspace: Settings → Context engine | Whether agents and the knowledge tool may compute over spreadsheets. |
| Sources and connectors | Project: Knowledge | Uploads and connected sources that feed the project. |
| Who can see what | Project: Knowledge | The audience of each source and each uploaded document. |
| Redaction | Project: Redaction | Rules that mask what the knowledge base returns. |
| Hybrid or graph | Each upload | Whether a file also builds the Context Graph. |
Managed or your own database
In Settings → Context engine, the first choice is where this workspace’s documents, embeddings and search run.
Managed by promptev
The default. promptev stores the knowledge and runs the models. Nothing to set up.
Your own database
Your documents, embeddings and Context Graph are stored in your own PostgreSQL database, and read with your own models and keys. Your data stays in accounts you control.
Going back to managed. Switching back disconnects the workspace’s database. The documents in it stay where they are, and the workspace runs on the managed engine at once. Connecting the database again later brings your setup back.
Your own database and models
With Your own database picked, the panel asks for the connection and the models. Every key and the connection URL are write-only: once saved, they are never shown again. Each part has a Test connection button, so you know it works before you save.
Database
The connection URL of your PostgreSQL database. The panel says whether the workspace is connected. When you save, promptev creates what it needs in that database, including the Context Graph when an extraction model is set; the panel lists what the database must allow.
Models
Your own database runs on your own models. Embeddings are required to connect; the rest are optional.
| Model | Needed | What it does |
|---|
| Embeddings | Required with your own database | Turns your documents and questions into vectors for search. Provider, model, dimension, an optional Base URL for a self-hosted or OpenAI-compatible endpoint, and your API key. The dimension must match the model and is fixed at the first upload. |
| Vision | Optional | Reads scanned pages and pictures when a file is uploaded. Without one, the extraction model reads them. The panel shows what reading 1,000 pages costs on the model you picked, charged by your provider. |
| Extraction model | Optional | Finds entities and relationships for the Context Graph. Add one to turn graph uploads on; hybrid search works without it. The graph is stored in your database, beside everything else. |
| Reranker | Optional | Puts the top search results in a better order before your agents read them. Cohere, Voyage, Jina or any Cohere-compatible endpoint, with your own key. If it ever fails, search keeps its own order. |
Model calls on your keys are billed by your provider. The models you set here take effect once the database is connected; until then the workspace runs on the managed engine.
Code execution
A switch in Settings → Context engine. It lets agents and the knowledge tool’s compute action run generated code over spreadsheets (CSV, XLSX), for sums, pivots and comparisons a model cannot do by reading. It is on by default. Turn it off for a workspace whose uploads you do not trust. When it is off, discover lists compute as off.
Sources and connectors
A project’s knowledge comes from its sources, on the project’s Knowledge page:
Uploads
Files uploaded in the app, from the SDKs or over the API all land here. Up to 20 files per call, 25 MB each. Uploading identical bytes again updates the document.
Connectors
Google Drive, SharePoint and OneDrive, Slack, Dropbox, Atlassian and more, kept in sync. A large set of files belongs in a connector source rather than in uploads.
A project API key reads and writes only its own project’s sources. source_ids in the knowledge tool narrows a call to some of them.
Who can see what
Each source, and each uploaded document, has an audience. You pick it when you add the source or upload the file, and you can change it later.
| Audience | Who sees answers drawn from it |
|---|
| Public | Anyone the agent serves, including anonymous visitors on embeds. |
| Everyone in this workspace | Any signed-in member. The default. |
| Selected people | Only the people (or, with SSO, the groups) you pick. |
Every search returns only what the person asking may see, and answers cite their sources. A project API key reads its whole project and nothing outside it.
Redaction
Each project can have redaction rules, on the project’s Redaction page. A rule is a pattern and the token that replaces each match. A preset adds common personal and structured identifiers in one click, and a test box shows what a rule masks before you save. The rules mask what the knowledge base returns, in every action and for every caller: the app, the SDKs, the API and MCP.
Formatted identifiers only. The preset catches identifiers with a known shape, such as dates, account numbers and IP addresses. Free-text names and addresses are not caught by it.
Hybrid or graph per upload
Each upload picks a mode. Hybrid, the default, makes the file searchable. Graph also extracts entities and relationships into the Context Graph, so questions about how things connect can be answered with traverse, find_related and get_neighbors. Spreadsheets are always read as hybrid. Graph is offered when the workspace has the Context Graph available; if it is not, the app shows “Graph (not configured)” and an admin adds an extraction model in Settings.
# Python: hybrid is the default; graph also builds the Context Graph
client.upload_document("contracts.pdf", mode="graph")
// TypeScript
await client.uploadDocument(files, { mode: "graph" });
# HTTP
curl -X POST https://api.promptev.ai/api/v1/documents \
-H "Authorization: Bearer fk_..." \
-F "files=@contracts.pdf" -F "mode=graph"An upload costs 1 credit per page, slide or MB (minimum 1). A search that reads the Context Graph costs more than a hybrid one; see the knowledge tool.