Retrieval infrastructure · EU data isolation
Search and chat over
your own documents.
You POST plain text. The API chunks it, embeds it, and gives you back semantic search, chat grounded in those documents, and an MCP endpoint. There is no vector database or embedding pipeline for you to run.
14-day trial · No credit card · EU data isolation
Query
how do refunds work for enterprise plans
-
refunds.md · chunk 2 0.89
Enterprise customers may request a refund within 30 days of initial subscription.
-
billing-faq.md · chunk 0 0.71
Refund requests are submitted through the account management portal.
-
terms.md · chunk 7 0.42
Subscription fees are billed monthly in advance and are non-transferable.
-
onboarding.md · chunk 1 0.19
Invite teammates from the team settings screen.
Anything scoring below 0.35 is dropped before it reaches your app.
Ingestion
What happens to a document you send
Every document goes through the same four steps. This is the pipeline you would otherwise build and keep running yourself.
-
01
Send text
POST a body, an identifier, and any metadata you want to filter on later. Send the same identifier again and the document is replaced in place.
-
02
Enrich and chunk
The text is converted to Markdown, summarised into problem and solution pairs, then split into 100–200 word chunks with configurable overlap.
-
03
Embed
Every chunk and pair is embedded with a managed embedding model. Identical text is embedded once, then reused everywhere it appears.
-
04
Retrieve
Queries are embedded, matched by similarity, reranked, and filtered. Your app receives documents and scores, not vectors.
The API
Store, search, and chat
Three calls cover the whole product. Every request takes a bearer token and a team token, and returns JSON.
curl -X POST https://api.vectorforgeai.com/v1/collections/abc123/documents \
-H "Authorization: Bearer $API_KEY" \
-H "Team-Token: $TEAM_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"identifier": "refunds",
"title": "Refund policy",
"body": "Enterprise customers may request a refund within...",
"metadata": { "department": "Finance" }
}'curl "https://api.vectorforgeai.com/v1/collections/abc123/documents/search?query=refunds&limit=3" \
-H "Authorization: Bearer $API_KEY" \
-H "Team-Token: $TEAM_TOKEN"
# Each document comes back with the similarity score
# that matched it, highest first.curl -X POST https://api.vectorforgeai.com/v1/messages \
-H "Authorization: Bearer $API_KEY" \
-H "Team-Token: $TEAM_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"user": "user_12345",
"message": "How do refunds work for enterprise plans?",
"collections": ["abc123"]
}'
# Returns a conversation_id. Send it back to continue
# the thread with its history intact.Model Context Protocol
Connect your collections to Claude or ChatGPT
The collections you already search through the API also work as an MCP server, so an assistant can read your documentation instead of someone pasting it into a prompt.
- Works with the collections you already have. No extra setup.
- One search credit per search_docs call.
- Authenticates with your existing API and team tokens.
claude mcp add --transport http "VectorForgeAI" \
"https://api.vectorforgeai.com/mcp" \
--header "Authorization: Bearer $API_TOKEN" \
--header "Team-Token: $TEAM_TOKEN"Use cases
What teams build with it
Support assistants
Answer customer questions from your help centre, with the source document attached to every answer. Ratings come back through the feedback endpoint so you can see what is working.
Search over internal docs
Handbooks, runbooks, and specs become searchable by meaning. Colleagues find the paragraph they need without knowing the file name.
Retrieval for your own model
Use search on its own and pass the chunks to whichever model you already run. The retrieval layer does not care what generates the answer.
Data handling
Where your data lives, and how long we keep it
The four answers a security questionnaire asks for first. The security overview has the rest.
- In transit
- TLS 1.3, falling back to TLS 1.2 for older clients.
- At rest
- AES-256, in ISO/IEC 27001 certified data centres.
- Conversations
- Deleted automatically after 30 days. Everything is removed when an account closes.
- Search terms
- Stored as SHA-256 hashes, so the query text is not kept alongside the vectors.
Pricing
Plans and pricing
Every plan starts with a 14-day free trial. Going over an allowance does not stop your app. The extra usage bills at the rates below, and you can set a cap.
Basic
For small teams starting out
$69 / month
- 300 documents
- 2,000 chat messages
- 50,000 search queries per month
- 1,000,000 LLM tokens in
- 500,000 LLM tokens out
- Analytics dashboard
- Community support
Pro
For growing businesses
$399 / month
- 800 documents
- 10,000 chat messages
- 500,000 search queries per month
- 5,000,000 LLM tokens in
- 2,500,000 LLM tokens out
- Advanced analytics dashboard
- Priority email support
- SLA
Enterprise
For large organisations
Custom
- Unlimited documents
- Unlimited chat messages
- Unlimited search queries
- Unlimited LLM tokens
- Advanced analytics dashboard
- Dedicated account manager
- SLA
- SAML and Okta
- SOC 2 report
If you go over
Overage is billed per unit at these rates. Set a budget in the dashboard to cap it.
- Documents
- $2.00
- per 100
- Search
- $0.50
- per 1,000 queries
- Messages
- $35.00
- per 1,000
- Tokens in
- $2.00
- per 1M · $8.00 on Pro models
- Tokens out
- $6.50
- per 1M · $26.00 on Pro models
What is VectorForgeAI?
An API that turns your text into a searchable, chat-ready knowledge base. You send documents, and you get back semantic search, conversations grounded in those documents, and an MCP endpoint. The storage, embedding, and retrieval run on our side.
What can I send to the API?
Text. Each document is a JSON request with an identifier, an optional title, the body as a string, and any metadata you want. There is no file upload endpoint yet, so you extract the text from PDFs or Word files before the call.
How is my text processed?
Each document is converted to Markdown, summarised into problem and solution pairs, then split into chunks of roughly 100 to 200 words. Overlap between chunks is configurable per collection and defaults to 17%. Every chunk is embedded and indexed.
Do I have to pick an embedding model?
No. Embedding is handled for you, and the model is chosen for the balance of retrieval quality and cost. You never call it directly, and you are never billed for it separately.
Which models answer my queries?
The Conversation API runs on a current frontier model and drops to a smaller one once a thread grows past roughly 75,000 tokens, so a long conversation stays affordable. The Responses API lets you choose a base or a Pro tier, priced differently. We keep these current as better models ship.
What happens when I pass a plan limit?
Your app keeps working. Usage above your allowance is billed at the overage rates listed in the pricing section. You can set a budget per usage type in the dashboard to cap what you spend.
Can I try it before paying?
Yes. Every plan starts with a 14-day free trial and no credit card is required to begin.
How is my data protected?
Data in transit uses TLS 1.3, and data at rest is encrypted with AES-256 in ISO/IEC 27001 certified data centres. Conversations are deleted after 30 days, search terms are stored as SHA-256 hashes, and closing an account removes everything. The security overview has the detail.
Send your first document today
Fourteen days, no credit card. The first document is one POST request.