Quickstart (CLI)

​

Overview

In this guide, we’ll develop an API to identify the top mentioned features in sales calls.

We’ll process sales call transcripts to extract insights like competitors mentioned, features mentioned, and whether it was successful. Our dataset, a 70-row CSV file, captures these activities.

Our steps include:

• Ingesting the dataset
• Creating AI columns
• Executing queries to filter, aggregate, and refine the into a top 2 list of features
• Exposing the top features through an API

​

Create a Workspace

Before we begin, we need to create a workspace to store our data and resources, as well as a token to authenticate our CLI.

1. Go to Airfold and create a new workspace.
2. Copy an Admin token from the workspace’s Token page.

The token should look like this: aft_6eab8fcd902e4cbfb63ba174469989cd.Ds1PME5dQsJKosKQWVcZiBSlRFBbmhzIocvHg8KQddV.

​

Set up the CLI

1. Install the CLI using pip install airfold-cli.
2. Run af config and paste your token when prompted.

pip install airfold-cli
af config

​

Create a Source

First, download sales_calls.csv

To create a source, a YAML file defining the source is required.

Let’s generate a source by inferring the schema from a CSV file
(replace /path/to/sales_calls.csv with actual filepath):

af source create path/to/sales_calls.csv .

The CLI will infer the schema from the CSV file and generate a YAML file with the following contents:

name: sales_calls
cols:
  ID: UUID
  Transcript: String
settings:
  engine: MergeTree()
  order_by: '`ID`'
  partition_by: tuple()

Push the source definition to your workspace:

af push ./sources/sales_calls.yaml

At this point, our source has been created but no data has been ingested yet.

Verify creation by listing sources:

af source ls

​

Ingest Data

With the source set up, ingest the CSV data:

af source append sales_calls path/to/sales_calls.csv

​

Create an AI Table

To derive insights, we can create AI columns that automatically extract competitor mentions, feature mentions, and success indicators.

In your workspace, create a analysis.yaml file with this configuration:

name: analysis
ai_cols:
  - name: successful
    type: Nullable(Bool)
    description: Whether this call was successful (which is defined as successfully
      setting up a follow up call). If the follow up scheduling is successful, the
      call should include an explicit date/time for the call.
    default_value: '""'
    using:
      - Transcript
  - name: competitors_mentioned
    type: Array(String)
    description: Who are the competitors mentioned (if any)?
    default_value: '"N/A"'
    using:
      - Transcript
  - name: features_mentioned
    type: Array(Enum('Customizable Templates', 'Inventory Management', 'Marketing
      Automation', 'Customer Support Integration', 'Social Media Integration', 'None'))
    description: Which features are mentioned in the call?
    default_value: '["None"]'
    using:
      - Transcript
cols:
  ID: UUID
  Transcript: String

ai_cols defines our AI-generated columns:

• successful: Identifies whether the call was successful based on the transcript content
• competitors_mentioned: Extracts any competitor names mentioned in the transcript
• features_mentioned: Lists any product features that were discussed

To see more details on the parameters of ai_cols, check out source.

Push this configuration to create the AI Table:

af push path/to/analysis.yaml

​

Ingest Data from sales_calls to analysis

Create a pipe to transfer data from sales_calls to analysis. Pipes in Airfold are used to run sequences of SQL queries that can transform, filter, and aggregate data across tables. Each step in a pipe is called a “node,” and each node runs a single SQL statement, allowing for more complex data analysis workflows.

Define it in a new pipe_to_analysis.yaml file:

name: pipe_to_analysis
nodes:
  - transform:
      sql: SELECT `ID`, `Transcript` FROM "sales_calls"
to: analysis

Push the pipe configuration:

af push path/to/pipe_to_analysis.yaml

Now, ingest data using:

af source append analysis path/to/sales_calls.csv

At this point, we have two populated tables:

• sales_calls: containing original data ID and Transcript
• analysis: containing data from sales_calls and additional AI columns

​

Further Analysis with Pipes

To identify top-mentioned features in successful calls, create an insights.yaml file:

nodes:
  - node1:
      sql: |
        WITH
          arrayJoin(features_mentioned) AS feature
        SELECT
          count() AS num_successful_calls,
          topK(2)(feature) AS top_features
        FROM
          analysis
        WHERE
          successful = true
publish: insights

• arrayJoin(features_mentioned) AS feature: This function “explodes” the features_mentioned array, creating a row for each feature mentioned
• count(): Counts the total successful calls
• topK(2)(feature): Extracts the two most frequently mentioned features
• WHERE successful = true: Filters the results to include only successful calls
• publish: publishes this as an endpoint, this line may be excluded if you do not wish to publish as an endpoint

Push the insights pipe:

af push path/to/insights.yaml

​

Query Results

Use the API:

curl --request GET \
--url https://api.airfold.co/v1/pipes/insights.json \
--header 'Authorization: Bearer <use auth token here>'

Or the CLI:

af pipe query insights --format json

Which should output:

{
  "meta": [
    {
      "name": "num_successful_calls",
      "type": "UInt64"
    },
    {
      "name": "top_features",
      "type": "Array(Enum8('Customizable Templates' = 1, 'Inventory Management' = 2, 'Marketing Automation' = 3, 'Customer Support Integration' = 4, 'Social Media Integration' = 5, 'None' = 6))"
    }
  ],
  "data": [
    {
      "num_successful_calls": "57",
      "top_features": [
        "Customizable Templates",
        "Inventory Management"
      ]
    }
  ],
  "rows": 1,
  "statistics": {
    "elapsed": 0.002417386,
    "rows_read": 70,
    "bytes_read": 2780
  }
}

​

Next Steps

You’ve successfully ingested, analyzed, and published data using Airfold in a few simple steps! This workflow enables intuitive interaction with unstructured data, transforming raw call transcripts into actionable insights.

Feel free to dive deeper into specific concepts, such as workspaces, sources, and more!