Choose the Right AI Model

Choose the Right AI Model

Feluda lets you choose which available AI model to use in Workbench and in your workflows.

Different models can produce different results from the same instruction. One model may be faster, another may follow complex instructions more closely, and another may be better suited to local or private work.

There is no single best model for every task.

The right choice depends on:

• what you want the model to do;
• how accurate or detailed the result needs to be;
• whether the task uses tools or attachments;
• where you want the information to be processed;
• how quickly you need a response; and
• what your computer can run when using a local model.

Start with the task

Choose the model based on the work you need to complete.

Ask yourself:

• Is this a short and simple task?
• Does it require detailed reasoning?
• Will the model work with a long document?
• Does the task need tools?
• Will it use an image, audio file, or another attachment?
• Does the result need a strict format?
• Will the same model be used inside a workflow?

A model that is useful for a quick summary may not be the best choice for a long comparison or a tool-assisted task.

Common task types

Use this table as a starting point.

Test the model with a real example before using it for regular work.

Cloud and local models

Feluda can work with supported cloud providers and compatible local model applications.

Cloud models

Cloud models run through an online AI provider.

They may be useful when you want:

• access to a broad range of models;
• strong performance on complex tasks;
• support for large inputs;
• tool or attachment features; or
• less dependence on your computer's hardware.

Using a cloud model normally requires an internet connection and an account with the selected provider.

Information sent to the model is handled by that provider. Review its terms and privacy information before using sensitive content.

Local models

Local models run on your own computer through compatible software such as Ollama or LM Studio.

They may be useful when you want:

• more processing to remain on your device;
• to work without an internet connection;
• direct control over downloaded models; or
• a model for repeated local tasks.

Local model quality and speed depend on the model and your computer.

A larger model may require more memory and may respond slowly on everyday hardware.

Choose based on privacy

Before choosing a model, consider the information it will receive.

For private or confidential work, check:

• whether the model is local or cloud-based;
• whether any tools connect to an external service;
• whether attachments are sent outside your computer;
• whether unnecessary personal details can be removed; and
• whether the provider's privacy terms fit your task.

A local model can keep model processing on your computer, but the complete task is only local when its tools, sources, and destinations are also local.

Do not assume that selecting a local model makes every connected action private.

Choose based on instruction following

A useful model should follow the structure of your instruction.

Test whether it can:

• include every required section;
• respect length limits;
• return a table or list correctly;
• use only the provided source;
• mark missing information clearly; and
• avoid adding unsupported details.

For example, ask the model to extract information using fixed headings:

Review the source below.

Return:
* Name
* Date
* Amount
* Required action

If a field is missing, write "Not provided."
Do not guess.

Compare the response with the source.

Choose based on accuracy

A model may write clearly while still making mistakes.

Accuracy matters when the task includes:

• names;
• dates;
• amounts;
• product details;
• policies;
• technical information; or
• decisions based on source documents.

Test the model with information you can verify.

Check whether it:

• preserves important details;
• separates facts from suggestions;
• identifies missing information;
• avoids invented claims; and
• reflects the source accurately.

Important results should always be reviewed by a person.

Choose based on speed

The fastest model is not always the best model, but speed can matter for simple or repeated tasks.

A faster model may be suitable for:

• basic classification;
• short summaries;
• simple rewriting;
• first drafts; and
• high-volume routine work.

A slower model may be acceptable when the task needs deeper analysis or a more careful response.

Compare speed only after confirming that the model produces an acceptable result.

Choose based on input size

Models differ in how much information they can handle in one request.

A short message is very different from:

• a long report;
• several documents;
• a large conversation;
• detailed instructions; or
• a workflow that adds information from tools.

When a model struggles with a long input:

• remove irrelevant information;
• divide the source into smaller sections;
• summarise earlier material first;
• start a new conversation; or
• test another model.

More input is not always better. Give the model only what it needs.

Choose based on tools

Some tasks require the model to use tools in Workbench or Studio.

Before selecting a model for a tool-assisted task, test whether it can:

• recognise when the tool is needed;
• choose the correct available tool;
• pass the required information;
• use the returned result accurately; and
• continue after a tool call.

Enable only the tool needed for the test.

Review the Activity log to confirm what the model actually did.

A model that produces strong text may not be the best model for a task that depends on reliable tool use.

Choose based on attachments

When a task uses an image, audio file, or document, confirm that the selected model supports that type of content.

Test whether it can:

• access the attachment;
• identify important details;
• follow your instruction;
• describe uncertain content honestly; and
• avoid guessing when information is unclear.

Compare the response with the original attachment.

Do not assume that every model available through a provider supports the same attachment types.

Choose a local model that fits your computer

Local models place different demands on your computer.

Smaller models usually:

• need less memory;
• start more quickly;
• respond faster;
• work better on everyday hardware; and
• handle simple tasks well.

Larger models may provide stronger results for some tasks, but they can need much more memory and processing power.

Begin with a smaller general-purpose model when you are unsure what your computer can run.

Move to a larger model only when:

• the smaller model does not meet the task;
• your computer has enough available memory; and
• the slower response is acceptable.

Check the model information shown by Ollama or LM Studio before downloading it.

Test models fairly

Use the same instruction and source information when comparing models.

A fair comparison follows these steps:

1. Choose one realistic task.
2. Prepare one clear instruction.
3. Use the same source for every model.
4. Start a new conversation for each test.
5. Keep tools and settings the same.
6. Review each result using the same checklist.

Do not improve the instruction for one model without repeating the revised test with the others.

Use a review checklist

Score each model using the needs of your task.

Choose the model that performs well on the points that matter most.

A practical comparison example

Imagine that you need to turn meeting notes into a structured update.

Use this instruction with each model:

Read the meeting notes below.

Return:
1. a summary of no more than 80 words;
2. a list of decisions;
3. a table with Owner, Action, and Deadline; and
4. a list of unanswered questions.

Use only information from the notes.
If an owner or deadline is missing, write "Not provided."

Meeting notes:
[Paste the same notes for every test.]

Compare the answers for accuracy, format, clarity, and speed.

The most useful model is the one that meets your actual requirements, not necessarily the one with the longest answer.

Use different models for different tasks

You do not need to use one model for everything.

You may choose:

• a fast model for simple classification;
• a stronger model for detailed analysis;
• a local model for private source material;
• a model with attachment support for images or audio; and
• a model with reliable tool support for connected actions.

Feluda lets you choose a provider and model in Workbench and in the relevant workflow steps.

This makes it possible to match each task to an appropriate model.

Use more than one model in a workflow

A workflow can use different models for different AI steps.

For example:

1. a local model classifies the input;
2. another model prepares a detailed analysis; and
3. a faster model formats the final result.

Use multiple models only when each one has a clear purpose.

More models can make a workflow harder to review and maintain.

Begin with one model. Add another only when testing shows a clear benefit.

Know when to change models

Consider another model when the current one repeatedly:

• ignores important instructions;
• produces the wrong structure;
• invents missing information;
• struggles with the source length;
• cannot use the required tool;
• does not support the attachment type;
• responds too slowly; or
• cannot run comfortably on your computer.

First confirm that the instruction and setup are correct.

A clearer instruction may solve the problem without changing models.

Save what you learn

Record which models work well for your common tasks.

Note:

• the task;
• the model;
• the instruction used;
• the strengths of the result;
• any repeated problems;
• tool or attachment support; and
• whether human review is required.

These notes make it easier to choose a model when building a workflow later.

Choose a model for Studio

Before selecting a model in a workflow:

1. test the task in Workbench;
2. use realistic sample input;
3. confirm that the model follows the required format;
4. test any required tools;
5. review errors and edge cases; and
6. confirm that the result can be checked.

A model should not be placed into a repeated workflow only because it answered one example correctly.

Test several different examples first.

Start with a practical choice

For a first model:

• choose a general-purpose model;
• use a simple, non-sensitive task;
• test one clear instruction;
• review the result against the source; and
• compare another model only when needed.

The goal is not to find one perfect model.

The goal is to choose a model that performs the current task clearly, accurately, and reliably enough for you to review and use.