Cortex - AI knowledge and operations platform
Cortex retrieves the right passage from your docs, tickets, and procedures, drafts an answer, and cites exactly where it came from - so a person can verify it before it reaches anyone. The first concrete, paid workflow is support: connect your knowledge base and Freshdesk, and Cortex drafts cited replies your team reviews before sending.
Also shipping software? See Canopy Host, our engineering infrastructure →
The problem
The answer to most repeated questions - a support ticket, a product question, an internal "how does this work" - already exists somewhere in your docs, tickets, or procedures. Finding it is manual work. General-purpose AI chatbots don't know your internal knowledge, and a confident wrong answer is worse than no answer at all.
Cortex solves this at the retrieval and drafting level: ingest your knowledge base, and Cortex retrieves relevant passages, drafts an answer, and cites its sources. A person verifies and sends. The first place this runs as a paid workflow is support - connect Freshdesk and Cortex drafts cited replies your team reviews before sending.
What Cortex is
What Cortex can do
Every generated message stores its citations as real records, not free text - each one points to the specific document chunk it came from and includes the exact snippet used, so "where did this come from" is answerable by looking at the message, not by trusting the model.
Upload PDFs, Markdown, HTML, and plain text files. Connect external sources via webhooks or scheduled sync. Cortex chunks, embeds, and indexes everything automatically.
In the Symphony support workflow, Cortex never sends a reply on its own - every draft is queued for a person to check, edit, and approve first.
Cortex combines vector similarity, Postgres full-text search, and trigram similarity with configurable RRF weights. Optional cross-encoder re-ranking is available behind configuration.
Create keys scoped to read, chat, ingest, or full access. Rotate one integration without affecting the rest.
Document uploads, assistant queries, API key usage, credential access, admin actions, and GDPR requests are logged per project.
Point existing OpenAI client code at your Cortex project endpoint and get retrieval-augmented answers from your own knowledge base.
A separate CLI and MCP server, scoped to Cortex’s own resources - not Canopy Host’s. Script knowledge-base and assistant setup from the CLI, or let an MCP-aware coding agent query your knowledge base as a tool.
For teams with specialised domain language once the retrieval and drafting workflow is proven. Not required to start a pilot.
How it works
In this workflow, Cortex never sends a reply on its own. It drafts. Your team reviews, verifies the sources, and sends.
A customer question comes in through Freshdesk (Zendesk next).
Cortex searches your trusted docs, tickets, and procedures for relevant context.
A draft answer is generated from the retrieved sources, not from memory.
The draft links back to the exact document and passage it came from.
A support agent reads the draft and the sources before doing anything else.
The agent edits if needed, then sends the reply from Freshdesk.
The ticket, sources, draft, and outcome are logged for review and improvement.
Prove it during the pilot
We do not claim results before you have measured them on your own tickets. The pilot exists to track these metrics on your data, so you decide with evidence, not a demo.
Track how long agents spend searching docs and old tickets before Cortex, and after.
Measure how quickly a draft is ready for review once a ticket arrives.
Track how often agents send a Cortex draft with little or no edit.
Measure the share of tickets where Cortex produces a usable starting point.
Track how often agents rewrite or discard a draft, and why.
Spot-check whether cited sources actually support the drafted answer.
Track whether fewer tickets need senior or specialist escalation.
Measure how quickly a new support agent becomes productive with Cortex drafting from day one.
The first packaged Cortex workflow
Symphony connects your Freshdesk workspace to Cortex. It reads incoming tickets, retrieves relevant context from your knowledge bases, and surfaces AI-drafted replies with citations your support team reviews before sending. It is not a separate product - it is Cortex, packaged for the support use case.
Who uses Cortex
Problem: Support agents search through documentation, old tickets, and procedures to answer repeated customer questions.
With Cortex: Cortex drafts answers from your knowledge base with citations the agent reviews before replying. This is the first Cortex workflow.
Problem: Product knowledge is scattered across Notion, Confluence, Jira, repo docs, and changelogs.
With Cortex: Ingest the material and ask questions about features, decisions, and specs with source-backed answers.
Problem: You need AI in your product, but building a reliable RAG stack takes months.
With Cortex: Use the OpenAI-compatible endpoint with scoped API keys, streaming, webhooks, and no shared model training.
For developers
If your application already uses OpenAI, you can point it at Cortex and keep the same client pattern. Retrieval, citations, and project isolation happen behind the endpoint. Cortex also has its own CLI and MCP server - separate from Canopy Host's - for scripting knowledge-base setup or letting an MCP-aware coding agent query your knowledge base directly.
from openai import OpenAI
client = OpenAI(
base_url="https://api.cortex.canopy.pm/v1",
api_key="crtx_sk_...",
)
response = client.chat.completions.create(
model="cortex-assistant",
messages=[{"role": "user", "content": "What is the refund policy?"}],
)
# Citations are included in the responseModel providers
Cortex works with your existing OpenAI or Anthropic subscription. Your provider key is encrypted at rest with AES-GCM envelope encryption and is never returned in API responses or logs.
Data ownership
Every document, embedding, conversation, and API key in Cortex is scoped to your product instance. Your knowledge base data is never used to train shared or platform-level models.
Cortex support pilot
Once we have the access and data we need from you, we target a working pilot within 14 days. Pricing is locked in for the full 90-day pilot. You leave with real numbers: search time, draft acceptance rate, citation accuracy, and more.
FAQ
Yes. Cortex uses retrieval-augmented generation with hybrid retrieval: vector similarity, full-text search, and trigram matching before generation.
No. Cortex drafts a reply and attaches citations. A support agent reviews, edits if needed, and sends it. Cortex does not send on its own.
Symphony connects your Freshdesk workspace to Cortex over webhooks. Incoming tickets trigger retrieval against your knowledge base, and a draft reply with citations appears for agent review.
Freshdesk is the first integration. Zendesk is planned next - contact us if that is a blocker for your pilot.
No. Most teams use the default retrieval and drafting pipeline. Fine-tuning is optional, advanced, and only useful for specialised domain vocabulary later on.
You need a model provider API key. Cortex supports OpenAI and Anthropic. You bring your own key and pay that provider directly.
Cortex provides GDPR erasure and export endpoints, EU infrastructure, project isolation, and no shared model training. A DPA is available on request.
Project data is retained for 30 days after cancellation. You can export your data before deletion.
Book a pilot
We prioritise teams with a Freshdesk workspace, a real backlog of docs and old tickets, and a willingness to work closely with us for the first 90 days.