Concept samples: Instabase
Context
This concept writing sample, part of Instabase's AI Hub documentation, helps users understand choose among the available large language models based on use case, cost, and capability. This sample reflects my approach to clearly explaining complex product tradeoffs for a mixed technical audience.
Choosing a model
Choose the model that best suits your needs, balancing speed, accuracy, and cost. More advanced models are better at reasoning and return more accurate results, but they’re slower and more expensive to use.
Model availability
The models available for use vary across AI Hub products. Some products support different models when using the API compared to the user interface.
Automation projects offer the standard model and, for commercial and enterprise users, the advanced model. You can switch to the advanced model for individual fields in the field editor.
Conversations offers the advanced and multistep models, as well as research mode. For queries to a single document, the advanced model is automatically used. For queries to multiple documents, the multistep model is automatically used. Research mode can be manually enabled for any query. Research mode uses a more powerful but slower variant of the multistep model.
The converse with a document endpoint, used to query single documents in a conversation by API, supports the standard and advanced models.
The run query endpoint, used to query chatbots or multiple documents in a conversation by API, supports the multistep model and research mode.
Chatbots offer the advanced and multistep models, as well as research mode. For queries to a single-document knowledge base, the advanced model is automatically used. For queries to a multiple-document knowledge base, the multistep model is automatically used. Research mode can be manually enabled for queries to a single-document or multiple-document knowledge bases. Research mode uses a more powerful but slower variant of the multistep model.
The run query endpoint, used to query chatbots or multiple documents in a conversation by API, supports the multistep model and research mode.
The following table provides a summary of model availability across AI Hub products. Models incur differing usage costs, something to consider alongside model capabilities.
| Model | Automation projects | Conversations | Chatbots | Conversations by API | Chatbots by API |
|---|---|---|---|---|---|
| Standard | Yes | No | No | Yes | No |
| Advanced (Commercial and enterprise tiers) | Single-document queries (automatic selection) | Single-document knowledge bases (automatic selection) | Yes | No | No |
| Multistep | No | Multiple-document queries (automatic selection) | Multiple-document knowledge bases (automatic selection) | Yes | Yes |
| Research mode | No | Single- or multiple-document queries (manually enabled) | Single- or multiple-document knowledge bases (manually enabled) | Yes | Yes |
Model capabilities
All models use certain capabilities in developing answers to your prompts, but they vary in complexity, accuracy, and approach. See the following table for a summary and comparison.
Visual reasoning is supported in automation projects when using the visual reasoning field type, which requires the advanced model. Visual reasoning is supported in conversations and chatbots only when using research mode, not when using the advanced model.
Standard model
The standard model is good at quickly analyzing most documents. It’s optimized for speed and cost, but it can be less accurate when it comes to longer documents, complex reasoning, or mathematics.
Standard model capabilities include:
- Search — Searches all pages in the document for information relevant to the prompt.
- Calculate — Performs basic mathematical calculations based on information found in the document.
- Extract tables — Extracts information from tables found in the document.
Advanced model
The advanced model is better at handling dense or complex documents, reasoning, and mathematics. The advanced model tends to give more accurate and detailed results.
| Model capabilities | Standard | Advanced | Multistep | Research mode |
|---|---|---|---|---|
| Search | Yes | Yes | Yes | Yes |
| Calculations | Basic | Basic | Complex | Complex |
| Table extraction | Basic | Complex | Complex | Complex |
| Plan next step | No | No | Yes | Yes |
| Evaluate response | No | No | No | Yes |
| Visualizations | No | Yes | No | Yes |
| Visual reasoning | No | Yes, in automation projects | No | Yes, in conversations and chatbots |
In automation projects, the advanced model is available only to commercial and enterprise users. In conversations and chatbots, the advanced model is automatically used for queries to single documents or single-document knowledge bases.
Advanced model capabilities include:
- Search — Searches all pages in the document for information relevant to the prompt.
- Calculate — Performs mathematical calculations based on information found in the document. The advanced model is better at mathematics than the standard model.
- Extract tables — Extracts information from tables found in the document. The advanced model more accurately retrieves data from tables, especially long or multipage tables.
- Visualize — (conversations and chatbots only) Creates visualizations such as graphs and charts, if you request a graph or chart in the prompt.
- Visual reasoning — (Visual reasoning fields in automation projects only) Analyzes visual and stylistic elements, including elements that OCR doesn’t capture, such as images, diagrams, watermarks, layout, colors, text styling, and handwritten markup.
Multistep model and research mode
The multistep model and research mode are available in conversations and chatbots. For queries to multiple documents or multiple-document knowledge bases, the multistep model is automatically used. You can enable research mode for any query using the Research mode toggle in the chat box.
When you give the multistep model a prompt, the model responds dynamically: it breaks your query down into a series of steps, executing each step before generating the next step. This step-by-step approach allows the multistep model to perform more complex reasoning and mathematics, better understand context, and return more robust and detailed answers. The same applies when using the more powerful but slower variant of the multistep model that’s used for research mode.
Multistep model capabilities include:
- Search — Searches all pages in the document for information relevant to the prompt. The multistep model performs reasoning or calculations based on the information it finds. It also returns details about what information it found, and where in the document it found the information.
- Calculate — Performs mathematical calculations based on information found in the document. Multistep model calculations are more accurate than those of the standard or advanced models.
- Extract tables — Extracts information from tables found in the document.
- Plan — The model considers its findings so far before determining its next step, resulting in more complete contextual answers.
The variant of the multistep model used for research mode shares the same capabilities, along with additional ones that make it well-suited to complex reasoning or computation queries.
While in research mode, additional capabilities include:
- Evaluate — The model evaluates its potential response to make sure it’s accurate and answers the prompt. If needed, the model revises its response before returning it to you.
- Visualize — Creates visualizations such as graphs and charts, if you request a graph or chart in the prompt.
- Visual reasoning — Analyzes visual and stylistic elements, including elements that OCR doesn’t capture, such as images, diagrams, watermarks, layout, colors, text styling, and handwritten markup.