#Ollama-JS - Add to Primal-Terrain. Someday.

1 messages · Page 1 of 1 (latest)

keen moat
#

https://github.com/ollama/ollama-js

This is a sick library;
https://shadowquill.org/

is a future wrapper of this. Enabling a localized chat.

The only cost, the processing power of your device. Requirements are bare minimum.

MY NEXT STEPS
Ollama-JS

  • If we use something like this, confirm that we store the model, the -JS lib of this saw docs on /cloud
  • What are our transcription equivalents.
  • How does this fit into PrimalTerrain if it does/purpose?
GitHub
GitHub - ollama/ollama-js: Ollama JavaScript library

Ollama JavaScript library. Contribute to ollama/ollama-js development by creating an account on GitHub.

keen moat
#

First use place idea could be for "intelligence" > "agent" use case.

When you "Root Your Location" on /new creation flow. Municipality "RAW" transformed into defaults.

desert parcel
#

Nice I like where you are going

#

First focus on your hardware contraints. ie what is the hardware specs for below average to average users.

#

Once you have a good estimation of that, pick a model that can run on said hardware specs

keen moat
#

I think hardware constraints is 'self hosted' since I would leverage as an API that I host. So it would be my processing.

The model choice would be based on what it is I'm doing?

So maybe a 'low compute req' model would be best suited to process:

"RAW" context:

"context": {
                    "district": {
                        "mapbox_id": "dXJuOm1ieHBsYzpBV25tN0E",
                        "name": "Vernon County",
                        "wikidata_id": "Q501587"
                    },
                    "region": {
                        "mapbox_id": "dXJuOm1ieHBsYzpBNlRz",
                        "name": "Wisconsin",
                        "wikidata_id": "Q1537",
                        "region_code": "WI",
                        "region_code_full": "US-WI"
                    },
                    "country": {
                        "mapbox_id": "dXJuOm1ieHBsYzpJdXc",
                        "name": "United States",
                        "wikidata_id": "Q30",
                        "country_code": "US",
                        "country_code_alpha_3": "USA"
                    },
                    "place": {
                        "mapbox_id": "dXJuOm1ieHBsYzpGRXlvN0E",
                        "name": "Viroqua",
                        "wikidata_id": "Q1569775"
                    }
                }

into normalized JSON/my actual Municipality fields

#

^ that above context changes dependent upon what is searched; which makes manual normalization almost impossible; especially if you do 'global' level stuff

maybe not a great use case, but its the first I've thought of for OLLAMA.

desert parcel
#

If you wanna get that deep, you should probaby set up RAG vector storage

keen moat
#

🙂 The lens of software can be so infinite.

desert parcel
#

use postgres for the data

#

good to go

keen moat
desert parcel
#

go step further and just train ur own model 😁

keen moat
#

Ahhhhhh.

"Intelligence" / "Agents" is not my priority in this App at the start.

this seems like the "next" step/when I really upgrade that.

I really want to go here already, but i need my groundwork solid/the mapping aspect to be 90% first 🙂

#

Great info/thanks for the share btw-- I'll review as I can. This is cool to see the opensource route.

keen moat
#

Yeah; I think this is just very simple for devs to integrate as I dig into it more.

http://localhost:11434
exposes an API.

sample

 const response = await fetch('http://localhost:11434/api/generate', {
        method: 'POST',
        headers: {
          'Content-Type': 'application/json',
        },
        body: JSON.stringify({
          model: 'gemma3:27b',
          prompt: prompt,
          stream: false // Set to true for streaming responses
        }),
      });

If I eventually host this as a website. I'm not sure I save much with going Ollama vs containerizing/hosting this with the next.js app. That seems to be the "next" part.

keen moat
#

Current best path for if I go "public" with the application is to leverage an experience like this.

Initial release I really want transcription; having the models and processing offloaded to the user is GIGGANUTS.

https://huggingface.co/spaces/Xenova/whisper-webgpu