Building with Supercog - supercog-ai/community GitHub Wiki
Supercog helps you do work and automate tasks using the power of agents - LLM-based programs connected to live systems and data.
Connected Co-pilot
Supercog goes way beyond the built-in training of your LLM. By accessing live applications and data, and by learning about your environment, Supercog is able to help accelerate work in a much deeper way.
-
ChatGPT has read the docs about your favorite API, but how does that API work in practice? Supercog can call APIs directly, examine their behavior, and update its learning based on what it finds.
-
SC can also call those APIs on your behalf to give you an answer. Want to know how many issues were filed on your Github repo in the last week? Just ask SC to answer the question and it will use the Github API to find out. Don't know what API endpoint returns that info? - me neither!
-
Try exploring your systems and data. SC can examine customizations inside a SaaS app, or examine sample data records. SC also brings a wealth of domain knowledge to this exploration, so it can help explain what certain fields mean, or how certain tables are likely being used.
-
SC understands SQL data stores and lots of file formats. It can also synthesize system-specific queries like Salesforce SOQL and JIRA JQL.
-
Need to cross-check with some other info? Add a browser tool and let Supercog lookup new information (guide, Swagger docs, live info) from the internet.
Prototyping new AI Applications
Supercog is great for experimenting and for prototyping new generative AI applications. Building applications with code takes a lot of work, while prototyping with ChatGPT is limited to writing prompts and document analysis.
With Supercog you can prototype a whole application as an Agent, including multi-step flows, testing different LLMs, and mixing lots of different actions including:
- text generation and code generation
- document and image analysis
- pulling information from the web
- text to SQL
- file and data processing
- quickly accessing APIs
Processing data
Do you have any "data monkey" jobs to do? These come up all the time - the need to pull data from somewhere, manipulate it, push it to somewhere else. Supercog excels at these "glue" tasks. It has a smart "data processing" layer that allows the LLM to orchestrate functions that can query, transform, and insert thousands of records efficiently. Not only are basic data movement tasks easy:
please download all the Leads from Salesforce and save them to a Snowflake table
but data transformation and synchronization are similarly easy:
read the excel file new_leads.xlsx, then map the columns of the file to the columns from the Salesforce Lead object.
Include "LeadSource" field and set it to "Supercog".
Then upsert those records into Salesforce Lead objects matching on the "Email" field.
Data processing is enabled by using Pandas DataFrames to efficiently move data through the system.
Automating tasks
Beyond just interactive tasks, Agents can be configured to run autonomously in reaction to different events.
- Whenever a new user record appears in our app database, create an associated Contact record in Salesforce. Include web research results based on the domain of the email address.
- Every morning, summarize the updates to any JIRA tickets in a project and post the summary to Slack.
- Download new Load records from Salesforce, save them into a Google Sheet, and email the sheet to my SDR team.