Alright, my pals, I’m again with one other submit primarily based on my learnings and exploration of AI and the way it’ll match into our work as community engineers. In at present’s submit, I need to share the primary (of what is going to doubtless be many) “nerd knobs” that I feel all of us ought to concentrate on and the way they are going to impression our use of AI and AI instruments. I can already sense the joy within the room. In any case, there’s not a lot a community engineer likes greater than tweaking a nerd knob within the community to fine-tune efficiency. And that’s precisely what we’ll be doing right here. Positive-tuning our AI instruments to assist us be simpler.
First up, the requisite disclaimer or two.
- There are SO MANY nerd knobs in AI. (Shocker, I do know.) So, if you happen to all like this type of weblog submit, I’d be completely happy to return in different posts the place we take a look at different “knobs” and settings in AI and the way they work. Nicely, I’d be completely happy to return as soon as I perceive them, no less than. 🙂
- Altering any of the settings in your AI instruments can have dramatic results on outcomes. This contains growing the useful resource consumption of the AI mannequin, in addition to growing hallucinations and reducing the accuracy of the data that comes again out of your prompts. Take into account yourselves warned. As with all issues AI, go forth and discover and experiment. However achieve this in a secure, lab surroundings.
For at present’s experiment, I’m as soon as once more utilizing LMStudio working regionally on my laptop computer moderately than a public or cloud-hosted AI mannequin. For extra particulars on why I like LMStudio, take a look at my final weblog, Making a NetAI Playground for Agentic AI Experimentation.
Sufficient of the setup, let’s get into it!
The impression of working reminiscence measurement, a.okay.a. “context”
Let me set a scene for you.
You’re in the midst of troubleshooting a community difficulty. Somebody reported, or seen, instability at some extent in your community, and also you’ve been assigned the joyful activity of attending to the underside of it. You captured some logs and related debug data, and the time has come to undergo all of it to determine what it means. However you’ve additionally been utilizing AI instruments to be extra productive, 10x your work, impress your boss, you understand all of the issues which can be happening proper now.
So, you resolve to see if AI can assist you’re employed by way of the info sooner and get to the basis of the problem.
You hearth up your native AI assistant. (Sure, native—as a result of who is aware of what’s within the debug messages? Finest to maintain all of it secure in your laptop computer.)
You inform it what you’re as much as, and paste within the log messages.
After getting 120 or so strains of logs into the chat, you hit enter, kick up your ft, attain in your Arnold Palmer for a refreshing drink, and await the AI magic to occur. However earlier than you possibly can take a sip of that iced tea and lemonade goodnessyou see this has instantly popped up on the display screen:
Oh my.
“The AI has nothing to say.”!?! How might that be?
Did you discover a query so troublesome that AI can’t deal with it?
No, that’s not the issue. Take a look at the useful error message that LMStudio has kicked again:
“Attempting to maintain the primary 4994 tokens when context the overflows. Nevertheless, the mannequin is loaded with context size of solely 4096 tokens, which isn’t sufficient. Attempt to load the mannequin with a bigger context size, or present shorter enter.”
And we’ve gotten to the basis of this completely scripted storyline and demonstration. Each AI software on the market has a restrict to how a lot “working reminiscence” it has. The technical time period for this working reminiscence is “context size.” Should you attempt to ship extra information to an AI software than can match into the context size, you’ll hit this error, or one thing prefer it.
The error message signifies that the mannequin was “loaded with context size of solely 4096 tokens.” What’s a “token,” you marvel? Answering that might be a subject of a wholly completely different weblog submit, however for now, simply know that “tokens” are the unit of measurement for the context size. And the very first thing that’s executed whenever you ship a immediate to an AI software is that the immediate is transformed into “tokens”.
So what will we do? Nicely, the message offers us two attainable choices: we will improve the context size of the mannequin, or we will present shorter enter. Generally it isn’t a giant deal to supply shorter enter. However different occasions, like once we are coping with giant log recordsdata, that possibility isn’t sensible—the entire information is necessary.
Time to show the knob!
It’s that first possibility, to load the mannequin with a bigger context size, that’s our nerd knob. Let’s flip it.
From inside LMStudio, head over to “My Fashions” and click on to open up the configuration settings interface for the mannequin.
You’ll get an opportunity to view all of the knobs that AI fashions have. And as I discussed, there are loads of them.
However the one we care about proper now’s the Context Size. We are able to see that the default size for this mannequin is 4096 tokens. But it surely helps as much as 8192 tokens. Let’s max it out!
LMStudio offers a useful warning and possible motive for why the mannequin doesn’t default to the max. The context size takes reminiscence and sources. And elevating it to “a excessive worth” can impression efficiency and utilization. So if this mannequin had a max size of 40,960 tokens (the Qwen3 mannequin I exploit generally has that top of a max), you may not need to simply max it out instantly. As an alternative, improve it by a little bit at a time to seek out the candy spot: a context size sufficiently big for the job, however not outsized.
As community engineers, we’re used to fine-tuning knobs for timers, body sizes, and so many different issues. That is proper up our alley!
When you’ve up to date your context size, you’ll have to “Eject” and “Reload” the mannequin for the setting to take impact. However as soon as that’s executed, it’s time to benefit from the change we’ve made!
And take a look at that, with the bigger context window, the AI assistant was capable of undergo the logs and provides us a pleasant write-up about what they present.
I significantly just like the shade it threw my method: “…take into account searching for help from … a certified community engineer.” Nicely performed, AI. Nicely performed.
However bruised ego apart, we will proceed the AI assisted troubleshooting with one thing like this.
And we’re off to the races. We’ve been capable of leverage our AI assistant to:
- Course of a big quantity of log and debug information to determine attainable points
- Develop a timeline of the issue (that shall be tremendous helpful within the assist desk ticket and root trigger evaluation paperwork)
- Determine some subsequent steps we will do in our troubleshooting efforts.
All tales should finish…
And so you might have it, our first AI Nerd Knob—Context Size. Let’s evaluate what we realized:
- AI fashions have a “working reminiscence” that’s known as “context size.”
- Context Size is measured in “tokens.”
- Oftentimes occasions an AI mannequin will assist the next context size than the default setting.
- Rising the context size would require extra sources, so make modifications slowly, don’t simply max it out utterly.
Now, relying on what AI software you’re utilizing, chances are you’ll NOT have the ability to modify the context size. Should you’re utilizing a public AI like ChatGPT, Gemini, or Claude, the context size will depend upon the subscription and fashions you might have entry to. Nevertheless, there most positively IS a context size that can issue into how a lot “working reminiscence” the AI software has. And being conscious of that reality, and its impression on how you should utilize AI, is necessary. Even when the knob in query is behind a lock and key. 🙂
Should you loved this look underneath the hood of AI and wish to find out about extra choices, please let me know within the feedback: Do you might have a favourite “knob” you want to show? Share it with all of us. Till subsequent time!
PS… Should you’d prefer to study extra about utilizing LMStudio, my buddy Jason Belk put a free tutorial collectively referred to as Run Your Personal LLM Domestically For Free and with Ease that may get you began in a short time. Test it out!
Join Cisco U. | Be a part of theCisco Studying Community at present at no cost.
Study with Cisco
X | Threads | Fb | LinkedIn | Instagram | YouTube
Use #Ciscou and#CiscoCert to affix the dialog.
Learn subsequent:
Share:
