Enrich Dataset Rows with Dynamic Columns from the Dashboard

Save as support_tickets.csv:

ticket_id,customer_message,agent_response,response_metadata,priority
T001,My laptop battery drains in under 2 hours even after a factory reset.,I have opened a replacement request. You will receive a prepaid shipping label within 24 hours.,"{""resolution_time_hours"": 24, ""replacement_approved"": true, ""ticket_owner"": ""support_team""}",high
T002,I never received my order from three weeks ago. Tracking says it was delivered.,I see the delivery was marked complete. Let me file a lost shipment claim with the carrier right now.,"{""resolution_time_hours"": 48, ""replacement_approved"": false, ""ticket_owner"": ""logistics_team""}",high
T003,The app crashes every time I try to open my account settings.,This is a known issue in version 3.2. Please update to version 3.3 using the link below.,"{""resolution_time_hours"": 2, ""replacement_approved"": false, ""ticket_owner"": ""engineering_team""}",medium
T004,Can I change my subscription plan before the billing date?,Absolutely! You can change your plan at any time from Account > Billing. The new rate applies from the next billing cycle.,"{""resolution_time_hours"": 0, ""replacement_approved"": false, ""ticket_owner"": ""billing_team""}",low
T005,I was charged twice for the same order last month.,I have confirmed the duplicate charge and initiated a full refund. It will appear in 3-5 business days.,"{""resolution_time_hours"": 72, ""replacement_approved"": true, ""ticket_owner"": ""billing_team""}",high

1. Go to Dataset then Add Dataset
2. Select Upload a file (JSON, CSV) and choose support_tickets.csv
3. Name the dataset support_tickets and drop in the file

You should see a support_tickets dataset grid with 5 rows and the 5 uploaded columns.

1. Add Column then Dynamic Columns then Run Prompt
2. Name: summary
3. Model Type: LLM, Model: gpt-4o-mini
4. System: You are a support assistant that writes short, factual summaries.
5. User: Summarize this customer support message in one sentence: {{customer_message}}
6. Concurrency: 5
7. Test then Run

You should see a new summary column populated with a one-sentence summary for every row.

1. Add Column then Dynamic Columns then Classification
2. Name: sentiment, Column (source): customer_message
3. Add labels: Positive, Neutral, Negative
4. Choose a model, Concurrency: 5
5. Test then Run

You should see a sentiment column with one of the three labels assigned to each row. Labeling is model-dependent, so treat the exact assignments as illustrative rather than guaranteed.

1. Add Column then Dynamic Columns then Extract Entities
2. Name: entities, Column: customer_message
3. Instructions:

Extract: issue type, product, urgency level, and location (if present).
Return concise entity values.

4. Choose a model, Concurrency: 5
5. Test then Run

You should see an entities column listing the issue type, product, and urgency level parsed out of each customer message.

1. Add Column then Dynamic Columns then Retrieval
2. Select your Vector Database type

Pinecone:

Field	Value
Column	agent_response
Pinecone API Key	Your API key
Index Name	Your index
Namespace	Your namespace
Query Key	Your query key
Embedding Configuration (Type + Model)	Your embedding provider and model
Vector Length	Your index’s vector dimension
Number of chunks to fetch	3
Key to extract	text
Concurrency	5

Qdrant:

Field	Value
Column	agent_response
Qdrant API Key	Your API key
Qdrant URL	Your instance URL
Collection Name	Your collection
Embedding Configuration (Type + Model)	Your embedding provider and model
Vector Length	Your collection’s vector dimension
Number of chunks to fetch	3
Key to extract	text
Concurrency	5

Weaviate:

Field	Value
Column	agent_response
Weaviate API Key	Your API key
Weaviate Cluster URL	Your cluster URL
Collection Name	Your collection
Embedding Configuration (Type + Model)	Your embedding provider and model
Vector Length	Your collection’s vector dimension
Number of chunks to fetch	3
Search Type	Semantic Search or Hybrid
Key to extract	content
Concurrency	5

3. Test then Run

You should see a new column populated with the retrieved chunks for each row, pulled live from your vector database.

1. Add Column then Dynamic Columns then Extract a JSON Key
2. Name: resolution_time, Column: response_metadata
3. JSON Key: resolution_time_hours
4. Concurrency: 5
5. Run

If response_metadata is still typed as text (its type after a CSV upload), every cell in resolution_time comes back null, since there’s no JSON to extract a key from.

6. Open the response_metadata column header menu and select Edit Column Type, then set it to JSON
7. Re-run the resolution_time column

You should now see resolution_time holding just the resolution_time_hours value pulled out of each row’s response_metadata JSON, for example 24 for T001 and 48 for T002.

1. Add Column then Dynamic Columns then Conditional Node
2. Name: triage_output

Branch 1 (if):

• Condition: {{priority}} == "high" (references the priority column; verify the exact comparison syntax against your workspace before relying on it)
• Column Type: Run Prompt
• Concurrency: 5
• System: You are a senior support analyst.
• User: Write a detailed triage summary and next action for this high-priority ticket: {{customer_message}}

Branch 2 (else):

• Column Type: Run Prompt
• Concurrency: 5
• System: You are a support assistant.
• User: Write a one-line summary for this ticket: {{customer_message}}

3. Test then Run

You should see triage_output carry a detailed analyst summary for the high priority rows (T001, T002, T005) and a one-line summary for the rest.

1. Click Evaluate then Add Evaluations and select groundedness
2. Map: context to entities, output to agent_response
3. Add & Run

You should see a groundedness column report Pass or Fail for each row, with a Reason column explaining the verdict.