API knowledge sources: data from an external system as knowledge in the knowledge base
With an API knowledge source IntraGPT collects data from a system of your choice on a schedule and places it as knowledge in the knowledge base. Everybody can then chat about it without anything being queried at that moment. It reads only: the connection never writes back to the source.
What a knowledge source does for you
Collecting on a schedule
You record where the data comes from and how often it should be collected. From then on that happens by itself, with nobody having to press a button.
Making it available as knowledge
The collected result lands in your workspace's knowledge base and is made searchable there. Questions about it are therefore answered from your own knowledge, alongside your other sources.
Supplying an explanation
With every source you write in plain language how the data should be read. What a code means, which unit an amount sits in, which field to ignore. That is the difference between raw data and usable knowledge.
Keeping or overwriting
You choose whether each round becomes a new version or overwrites the previous one. For a current price list you want the latter; for something where you want to see change over time, the former.
Seeing whether it went well
Per source you see when it was last collected and whether that succeeded. If something goes wrong on the other side, it shows with the source instead of staying silent and going unnoticed.
Sources that lend themselves to this
Wholesale
- Situation
- Prices and terms live in a system not everybody is allowed to query.
- What the agent does
- The data is collected daily and made available as knowledge, with an explanation of the volume tiers.
- Result
- Sales can answer pricing questions without access to the source system and without calling the department.
Trade association
- Situation
- A public register changes now and then and staff keep looking it up by hand.
- What the agent does
- The register is collected periodically and added to the knowledge base, with an explanation of the fields.
- Result
- Questions get answered in the chat, with the same content as in the register itself.
Logistics
- Situation
- A planning system could not cope with an entire department querying it live.
- What the agent does
- The relevant data is collected a few times a day and placed in the knowledge base.
- Result
- Everyone can ask questions without the source system suffering for it.
Healthcare
- Situation
- A national list of guidelines is maintained centrally and has to be consultable internally.
- What the agent does
- The list is collected periodically and added as knowledge, so the assistant can quote from it.
- Result
- Staff get answers from the current list rather than from a copy somebody once saved.
Adding a source
-
1
Recording the source
You give the source a recognisable name and record where the data comes from and how access is granted. That can be a key or a username and password, depending on what the other side asks for.
-
2
Choosing the frequency
You decide how often collection happens. For a list changing daily, once a day is plenty; for something changing more often you pick a shorter round. A source can be paused with a single switch.
-
3
Adding the explanation
You write down how the data should be read. This is the step that pays off most and gets skipped most often; take your time over it, because it decides the quality of every answer.
-
4
Following and cleaning up
Per source you see the last round and whether it succeeded. If you remove a source, the stored data goes with it and the knowledge base is searched anew, so no remnants linger.
What comes in, and what never goes out
This connection only reads. That keeps the security question manageable: it is about what comes in and who may see it, not about what could change in an external system.
Where the data lives: Nederland of de EU
One way traffic
There is no way to write back to the source. What you set up here can collect data and nothing else, not even when somebody asks for it.
Visible to whoever you choose
The collected knowledge falls under the same agreements as the rest of the knowledge base. You decide which assistants and which roles may reach it, so not everything has to be available to everyone.
Removing really removes
If you delete a source, the collected data disappears and the knowledge base is updated. No copy lingers that could surface in an answer later.
Storage and processing with us
The collected data sits in your workspace's environment and is processed on our own server in the Netherlands. It does not go to an American model vendor and is not used for training.
Questions about API knowledge sources
What is the difference with an ordinary connector?
An ordinary connector is called during a conversation: you ask a question and the assistant goes and looks right then. A knowledge source collects in advance and puts the result in the knowledge base. That is nicer when the source is slow, has a limit or changes only a few times a day, and it keeps the source system calm.
Does IntraGPT work with our own data service?
Yes, as long as it is reachable over the internet and has some form of access control. You record the source once, choose how often it is collected and write down how the data should be read.
How current are the answers then?
As current as the collection round you choose. Pick a daily round and the answer is today's. If you need something correct to the second, an ordinary connector is the better choice, because that one looks live.
Can I connect this to ChatGPT?
ChatGPT cannot reach your internal sources and has no knowledge base that belongs to you. Here the data lands in your own workspace, with your permissions around it, and it is not used for training.
What happens when the source does not respond once?
That single round fails and you see it with the source, with the error message attached. Knowledge collected earlier simply stays, so nothing disappears from the knowledge base because a system was briefly offline.
Why does the assistant give a strange answer about this data?
Usually because there is no explanation with the source. Without it the model does not know what a code means or which unit an amount sits in. A few lines of explanation with the source almost always resolves that in practice.
What does an API knowledge source cost?
Setting it up is part of the track we go through together. What drives the work is the number of sources, how often they are collected and how much explanation the data needs. Thirty minutes is enough to know where we stand.
How long does setup take?
Adding a source takes minutes. The work sits in deciding which data you genuinely need and in writing the explanation alongside it. We prefer to do that last part together, because that is where the quality of the answers comes from.