Blog
Guides, quickstarts and product news — AI agents, RAG, workflows and MCP.
No evidence isn't “made it up”: three verdicts for an LLM judge
If you measure hallucinations with a model that grades past conversations, a gap in your knowledge base reads as the bot lying. How to separate “made it up” from “nothing to check it against”, and how to prove the rule holds.
Before you delegate a plan, ask git if it's already done
A plan describes the day it was written; a merged PR describes today's code. A two-minute check before handing out phases keeps you from sending someone (or an agent) to build what already exists.
A long chat's real cost is the history you resend on every turn
Every turn of an AI chat sends the whole conversation to the model again. Why that grows much faster than the chat itself, and how a window with a summary cuts it to a fifth without the user noticing.
One API key for one assistant: what it sees and what it doesn't
You hand a key to a freelancer or an agent so they can work on one assistant. What it reaches in CognitionDesk —assistants, channels, widgets, child keys— and how to check it with three requests.
Publish, list versions and roll back: an assistant's lifecycle from the MCP
In CognitionDesk the draft and the published version are two different things. Which one each channel runs, how to see what's live before you touch anything, and how to undo a bad publish without losing your edits.
Dense, hybrid and reranked: how much each search finds on the same knowledge base
We measured five RAG search modes on 48 questions, with confidence intervals. Before you add a reranker, look at your similarity threshold.
Freeze your API's shape, not its data
A value snapshot gets regenerated without a second look every time a date changes. A shape snapshot only moves when you break a client. How to build one, how to test it, and where the version goes.
Your SSE stream dies while the model thinks: how to keep it alive
If the model takes longer than your server's timeout to produce its first word, the stream dies before it starts. A three-step recipe and how to test it with a fake slow model.
Many clients, one MCP server: the account keyring
If you run several clients' CognitionDesk accounts, you don't need one MCP server per account. One variable with aliases, one tool to switch, and a line at the top of every response that tells you where you are.
The 117 tools in the CognitionDesk MCP, by area: what your agent can do with your account
You connected the CognitionDesk MCP. Here are its 117 tools grouped by what you want to do, which ones only read, which ones delete, and why nobody writes the table by hand.
What your agent can do and who it's talking to: two permissions
The tool's permission and the permission of whoever is writing are not the same thing. How to separate them with four rungs, a floor derived from the data and a closed default.
Add an AI chat agent to your website in 3 lines of code
Embed a RAG-powered AI agent that knows your business docs using the CognitionDesk widget — no backend required.