Inspecting Responses
ChainForge includes many features to help you inspect and navigate LLM outputs. To inspect responses once they've been collected, click on the footer of a node:
This brings up a full-screen inspector window with two views, described below. You may also:
- click the [> toggle in the bottom-right corner to use a pop-out inspector drawer (NEW)
- connect an Inspect Node for examining responses without a pop-up
The remainder of this document concerns the full-screen response inspector.
Grouped List Layout
Grouped list layout allows you to order responses into hierarchical groupings of prompt variables and LLMs. If you have queried multiple LLMs, responses will appear side-by-side for easy comparison across models, for up to five models. Here we view responses for two models, GPT3.5 and GPT-4:
Click the groups to expand or collapse them (by default, all but the first of these groups are collapsed):
You can also rearrange the order of the hierarchical grouping by prompt variables. For instance,
here there are two prompt variables command and input. We first group by command, then prompt:
Table Layout
The Table Layout is an alternative layout that may help for your use case. We've found it especially useful to compare across prompt templates by plotting prompt variables as columns.
By default, Table Layout plots LLMs as columns, with prompt variables in blue on the left:
You can change this to plot prompt variables as columns.
For instance, here we specify columns as different commands to professionalize an email.
Our example input, email, appears on the left-hand side:
Export Data
Click 'Export' on the top-right corner of an Inspector window to generate an Excel xlsx file with the data:
Prompt variables, LLM nicknames, prompts, responses, and any evaluation scores are included.