Improvement suggestion for constants for API and model usage

[DirtyHamster]

I was messing around with your project the other day looking to try to run it locally. I noticed the server address is buried at the end of autoscript.rpy whereas the rest of the constants for API and model usage are listed clearly at the top of main.rpy. If you were to move the server address to be a constant listed in main.rpy it will make it more consistent and a little easier to run this locally if one chooses to do so.

If you don't make that change you can still append the instructions to show people, they can run this locally and to go look for the following lines to edit in autoscript.rpy (bottom of file):

data['messages'].extend([{'role': 'user', 'content': message} for message in self.conversation_history])
try:
response = requests.post('http://12.0.0.1:1234/v1/chat/completions', headers=headers, json=data)
response.raise_for_status()
return response.json()['choices'][0]['message']['content']
except requests.HTTPError as http_err:
print(f"HTTP error occurred: {http_err}")
self.conversation_history.pop()
narrator("An error occurred while communicating with the GPT model. Please try again.")
return ""
except Exception as err:
self.conversation_history.pop()
print(f"Other error occurred: {err}")
return ""

I tested it on LM studio but I'm fairly sure it would work with others too.

[Wendy-Nam]

Thanks a lot for pointing this out! 🙏

You're absolutely right — it makes much more sense to keep the server address consistent with the other constants in main.rpy.

I'll update the project shortly to move it there and also add a quick note in the docs for those who want to run it locally.

Really appreciate you taking the time to test it out and share the feedback!

[Wendy-Nam]

By the way, could you also share which versions of LM Studio, Python, the openai (or requests) module, and Ren'Py you used for testing?
That will help ensure the instructions and changes are consistent across environments.

[DirtyHamster]

More than welcome, it's an interesting project. I might have some more ideas for you later. I still have to play around with it a bit more and get use to its quirks.

Current version of Renpy: 8.4.1

Current version of LM studio is 0.3.3.2 (build 3) However I think as long as it uses the GPT end points it should work (photo below of end points). LM studio is an all-in-one package for AI backend, so it does its own thing with python requirements probably within a virtual environment, The Link for them: https://lmstudio.ai/

Out of python and conda that I have on locally: I'm fairly sure none of these installed with LM studio, just documenting them in case.
I wouldn't even have standard python on if I didn't need it for pyinstaller and a few other programs.

The only extra instructions from me would be to start the LM studio server via its development mode: https://lmstudio.ai/docs/app/api and leave the api key empty in "your settings" as it doesn't require one server side yet.

Photo note: "end points to chatgpt": "So, it's as simple as swapping out the local address portion in the http addy"

I made sure just to match the end of what you had as it was in the list.

Outside of that the model I tested it on was: https://huggingface.co/mradermacher/dolphin-2.8-mistral-7b-v02-GGUF it's fairly good at uncensored storytelling and roleplaying.

There should probably be some or a note on token usage too the rough equivalent to 1 token is anywhere from .6 to 1mb of memory (kind of works as rough math but it's really not accurate). I left mine at default settings, but I didn't play a long game, so the limited token count didn't cause any problems for me yet. There's a token setting when you load the model.

[DirtyHamster]

in main I just listed the model's name as so:
MODEL_NAME = 'dolphin-2.8-mistral-7b-v02' # replace it with your model name

In LM studio they do list it differently as: mradermacher/dolphin-2.8-mistral-7b-v02

I didn't run into any issue though, but it might be pointed out in case something changes later. They do have it as a copy-able field in their gui though.

Outside of this and the last post I left everything default just to test it on your example story.

[Wendy-Nam]

Thanks a lot for your kind words — it really means a lot.
I actually made this project as a small experimental side-thing while practicing Python, and honestly I didn’t expect anyone to notice or use it now. That’s why your feedback and suggestions were such a pleasant surprise for me.

I’d love to try out LM Studio and also explore HuggingFace’s API integration — both sound really interesting. That said, I’m currently in the middle of job hunting and just got a rejection after a long process today, so it might take me a few weeks before I can properly improve and release new versions.

Still, your message gave me the motivation I needed. I’ll start with easier tasks like versioning and documentation, and later refine the architecture (callbacks + displayables to avoid Ren’Py runtime conflicts, etc.).

So once again, thank you — today was a tough day for me, but your encouragement really made a difference.

[Wendy-Nam]

@DirtyHamster

By the way, I wanted to ask for your opinion on something.

When I originally developed this project, I wasn’t yet familiar with using callbacks or similar techniques to prevent Ren’Py from freezing during AI response wait times. As a result, there might be short freezes while the model is generating.

In my more recent Ren’Py tool (“layouttools”), I implemented callbacks to avoid that issue — but since this project was older, it doesn’t have that improvement yet.

Have you found the freezing during response waits to be a major inconvenience? Or have you already found your own workaround while testing?

[DirtyHamster]

I did notice some, but it was less time than when I tried to have python and the AI reorganize my bookmarks (It did not work out well as far as organizing goes, it did do some work though as was much less organized then anticipated) or any other generative task that I've given. It's not something that really bothers me as I come from doing audio and video and am rather used to waiting for files to finish rendering. It might however bother or spook the general gamer that is not used to it or not expecting it.

I probably would try to come up with some type of a "generating content wait screen" to indicate the process when it's passing back and forth its information stream. I would probably do this as a character rest screen where you can look at your stat's and interact with the character, partner, or others before taking any further actions. I.e a loop return spot before next action is decided.

As per the is their work arounds, there are some, and I haven't tried any of them yet with your code. You could have the game preprocess areas before the character moves to them. I.e. get close enough to where the next content is needed and that auto triggers the next set of generation before you even move to it open world games often work like this, where your small sphere of influence kind of proceeds and retracts around the character. I would probably take a deep look at procedurally generated content as this can help guide the game as well as cut down on processing time and then bulk load an explorative area. I would also look at retaining areas rather than allowing them to vanish instantly. If you can retain an area you don't have to regenerate everything again however you can then play with area events and internal content like does a pigeon move or do items on the ground change.

When I played your demo, I noticed quickly that there really wasn't any area management concept involved as well as some other non-character-based limitations. It's not a horrible thing but it could hinder somethings like saving states and resuming where you left off. Most of the text based rpg games that I use to play were using databases to drive their worlds. Just because you want to generate the areas via ai automatically doesn't mean you can't also use preservation methods to reuse them.

I would also look at data structures for NPC's and MOBs more not being able to predefine these is going to make any type of narrative driven story harder to do as much as it is autoscripted there is probably narrative waypoints that could be used to help push the story along. If you'd like I could show you some that are historic via game history a lot of it will be the contrasts between rpg and narrative story telling though.

[Wendy-Nam]

I found your points very interesting. I had also thought about implementing a wait screen in the past, but at the time I didn’t know enough about Ren’Py’s Python structure or possible workarounds, so it felt difficult to attempt. As for the database, in the current version I’m only handling it by periodically backing up the game’s progress into JSON.

Since you mentioned there are historic cases worth benchmarking, I’m very curious to learn more. If you could share them, it would be a great help to me.

[DirtyHamster]

I get kind of lost in code too but I'm not much of a coder as much as I can read though a code base and understand it. I just never picked one to really learn. So, I can change things in code but not really write things from scratch. You should also look at Yaml files for some of the saves. It's easy to read and not everything needs to be transmitted to the AI which the Json files are better at doing. I'll often do my setting for gui in Yaml.

I can give you a few samples, 2 early ones that I would point at that can be still played online that often get brought up when talking about text-based gaming history: Zork 1 and 2
https://classicreload.com/zork-i.html
https://playclassic.games/games/adventure-dos-games-online/play-zork-ii-wizard-frobozz-online/

These 2 were not the easiest games you really should have some paper and a pen to make maps when you play them. I'm serious you'll need them. There are more games in this category, but they are often really limited in agency for the player. It is through a really good starting point for anything forward of it. As open agentic action for the user character is important, I'll take a really large leap to this.

I'm skipping a lot of multichoice adventure games as they are all hard scripted.

BBS door games and Muds there are still some running the few I would point out were cookie cutter systems that ran out of the box the same every time, Majormud, another that was rather enjoyable was called Legend of the red dragon (LORD). I'm pointing these out mostly because they were super consistent, there were lots of other text muds that were way more customizable. A lot of them are still playable online. Even though these are multiplayer they still have the same fundamentals as single player games within the spectrum as rpg adventure games.

Some info from Majormud:
Game owner site: https://www.majormud.com/main.html
https://www.mudinfo.net/index.php (better source for info on the game) Syntax is the person you should really talk to as he created a lot of the editors for the game.

Note: the game is grindy and most players use scripts to do the boring parts. i.e they make a script that plays the game for them for hours on end.

MMud was a heavy databased game it was often hacked/edited even though it was a cookie cutter game. The game has been kind of borderline abandonware for a number of years though. The only bad portion about the game is its economics when the game has been running for a long time. Starting a character on a newly started game is very different from starting on an older game as there are fewer valuable items tossed around like trash however the game play experience was generally the same on every system. A number of years ago a past friend recoded the game from scratch using C# and I did a lot of the bug testing on the project So this game in particular would be easy for me to tell you about.

One of the things I would point out with this one is the room descriptions could have been vastly simplified using some procedural generation. It probably could have been updated using structured data to form the room descriptions instead of repeating a lot of the text in the database for maps and rooms for every room. Take a look at this site: https://roomai.com/interior-design-glossary/hierarchy I would also look at: https://homelyville.com/room-types/ In a system like you have you could use these as blueprints for guidance giving the ai a clearer idea of what you're asking for. I would probably try to make some type of framework around points of interest and mapping that is reusable between games. I would look at what Zillow or other real state places list for what can go in and work around that. Provided there are samples the AI should be able to generate vs the pool of information provided. Even if you have multiple directions of movement the wording of the next room outside of a gateway to another area should be similar. Ease of editing really drives creation of new experiences. At more than one point I started doubting how mapping worked in the game, thinking it would be better to use way points and points of interest rather than roaming hard-set maps. I started asking myself stuff like what's the difference between house x and y if the variables were just customary. If the architecture is the same, then it's just the interior dressings that change.

Some of the interesting concepts out the programing in the game was energy usage per action and how turns were dealt with. you had a max amount of energy, and each action would subtract from that amount when you got to an unusable amount it then would give the turn to the next actor in which ever action engagement you were in.

Magic in the game worked similar each spell was databased and used a bit of energy till the turn was over, if it didn't use all the energy in a single casting it would repeat till energy was depleted. This type of construct is kind of important to have a stable artificially created world. Without them the AI can go really wild and suddenly summon donkey kong to drop a barrel out of nowhere. Each spell had an algorithmic structure to it to define its growth cycle against the characters stats,

Weapons and combat similarly used energy, but they were more dependent on character stats most of the calculations revolved around class combat inherited skill 1 to 4, agility and strength and character condition. Which was a classic D&D approach, I would suggest reading some of their rule books going forward, you don't have to follow them but there are useful ideas with how to deal with situations in gaming in them.

There were automatic player maintenance systems coded in for tracking and updating health and mana. I.e. at a wait state you would recover at one pace, at a rest state you would recover faster. While there was code in for survival aspects like thirst and hunger, they never implemented them in the content. How a character recovers is something you could add in, while playing the demo I did try to heal myself and the game agreed to do so even though I assumed I had zero inventory on hand which is a bit of a logical flaw.

Skill testing would be another thing I would look at trying to hard code a bit more. If I try to jump over something or dodge out of the way the game should look at the character stats (by stats I do mean both attributes and status conditions).

This is something you could incorporate as character management and the rest screen. I.e. if no action is needed a player can rest and heal.

In this game NPC's were hard coded as well as heavily scripted for context. I would look to make this more open i.e. local area knowledge, individual knowledge, current event knowledge (access to game action logs) .... Even if an NPC is just there to give a way point giving them some agency does vastly improve them. You should be able to state them as a character at a point of interest, they should be treated similar to a player character as well. (I think you could reuse your code from the partner but make it not follow for this. It would have to be defined by the game editor rather than player though too.)

Lord:
https://legendreddragon.net/
Lord used an instance type of exploration where you go to known areas but in those areas a set of situations could happen. This is probably closer to what you have right now in play style limited choices however you're missing any hardcoded points of interest that the character can visit. This game had a forest, an inn, a weapons shop and a few other known places that you could always go to. In these known areas there were lists of actions that would always be there unless some other status issue limited them. As much as I use to like this game there wasn't much to do.

Unlike the other this game is faster and easier to establish your character in. Some things I would point out is the point of interest this seems to be consistent between the two. While the game still ran on character stats that were hard coded it used them differently. You didn't wonder in maps you just went to locations this removed the issue of having a large database of rooms, but they still had known areas with known effects in those rooms. (I can clarify more if needed).

Some other programs I would point out to you as they deal with randomly generating content in a controlled way. These in some ways are more inspiring than the actual games.

https://www.mysticwaffle.com/maps (graphical but could easily be translated into text rooms)
https://donjon.bin.sh/fantasy/town/
https://www.thievesguild.cc/city/
https://www.vondy.com/random-city-map-generator--o7lkw76q
https://roll20.net/ (this is a full system)
https://www.redblobgames.com/maps/mapgen4/ (this is graphical, but a map is sometimes handy)
https://rollforfantasy.com/tools/population-generator.php (no city runs by itself think about nurse jenny from Pokémon you can reuse content. a temple healer can always be the same and can differ if the system looks at their locality.)

I have to admit I was a little lost on how exactly to answer you on this as there is a vast amount of info out there, but you can pick my brain if you want to for more details, I am really curious too on what practices work best when incorporating generative ai into games including what is useful from other systems.

There are also many more muds than I could cover in this short answer look around as some of them have really cool custom features.

[DirtyHamster]

I often look at mud as foundations for gaming I would look at some of the top lists for ones that are currently active: https://www.topmudsites.com/

Then I just start analyzing the systems they have in play for both player and npc/areas.

[DirtyHamster]

I'm also not sure how far you want to take various things in games like if you want a trading game, I would point you to different ones like:
http://www.pocxxl.com/overview/overview_en.html (uses more gui then text but could be a text game actually.)

or

https://tradewars.com/default.html (pure text)

[DirtyHamster]

Some other text based muds to look at: (a lot of these have been running for years.)
https://www.mudconnect.com/cgi-bin/search.cgi?mode=mobile_biglist

a fair list of current single player text games:
https://textadventures.co.uk/

This list also is a good source for some of the historical ones:
https://en.wikipedia.org/wiki/List_of_text-based_computer_games