feat(feedback): implement feedback summaries asset and job

[rachellougee]

What are the relevant tickets?

Close #2542

Description (What does it do?)

Builds the feedback_summaries Dagster asset

• Reads int__feedback__conversation, calls an LLM to generate conversation_summary per multi-turn conversation
• Skip rule: single-turn conversations, and multi-turn ones under 500 characters, are not summarized
• Incremental, and change-awareness: only summarizes new or changed conversations (a ticket that gains a new comment gets re-summarized, not skipped forever).
• Supports Anthropic, OpenAI, and AWS Bedrock as the LLM backend (needs to confirm with @blarghmatey if Bedrock is configured on production)
• Fixes a PII-redaction bug where a URL could be misread and redacted as a person's name.

How can this be tested?

int__feedback__conversation must exist in your dev schema. If it doesn't, copy it from production:

  create table ol_data_lake_production.ol_warehouse_production_<suffix>_intermediate.int__feedback__conversation                                                 
  as                                                                                                                                                             
  select * from ol_data_lake_production.ol_warehouse_production_intermediate.int__feedback__conversation;  

• Run the unit test suite:

 cd dg_projects/ml                                                                                                                                              
 uv run pytest ml_tests/ -q 

• Run the asset locally with your Parley key at https://platform.parley.mit.edu/my-keys

  export DBT_SCHEMA_SUFFIX=<your schema suffix>                                                                                                                  
  export DAGSTER_ENV=dev                                                                                                                           
  export ANTHROPIC_API_KEY=...                                                                                                                                   
  export ANTHROPIC_BASE_URL=https://parley.api.mit.edu

  uv run dagster dev -m ml.definitions    

• In the Dagster UI, materialize feedback_summaries with a small sample_limit via Launchpad config:

 ops:                                                                                                                                                           
   feedback_summaries:                                                                                                                                          
     config:                                                                                                                                                    
       sample_limit: 100      

Expected log line: Summarized 100 new conversations (0 already summarized, 100 total upstream).

• (Optional) Turn on the automation sensor at http://127.0.0.1:3000/locations/ml.definitions/sensors/feedback_summaries_automation_sensor.
  The sensor is off by default and this is only needed once Bedrock/API access is confirmed working and you want it to run automatically.

Query the output table

select * from ol_data_lake_production."ol_warehouse_production_rlougee_intermediate".feedback_summaries

Additional Context

For deployment step, do a Full refresh — feedback_redacted to fixes historical mis-redactions:

  ops:                                                                                                                                                           
    feedback_redacted:                                                                                                                                           
      config:                                                                                                                                                    
        full_refresh: true      

[Copilot]

Pull request overview

Adds an incremental Dagster asset for LLM-generated feedback conversation summaries, including provider support and PII-redaction improvements.

Changes:

• Adds conversation summarization, skip rules, and incremental processing.
• Adds Bedrock authentication and automation/job configuration.
• Updates PII redaction behavior and unit tests.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
ml/resources/llm.py	Adds Bedrock and local environment authentication.
ml/lib/summarize.py	Implements summarization and incremental filtering.
ml/lib/redact.py	Adjusts overlapping URL/PII handling.
ml/definitions.py	Registers the asset, job, sensor, and resource.
ml/assets/feedback_summaries.py	Defines the summaries Dagster asset.
ml_tests/test_summarize.py	Tests summary rules and incremental filtering.
ml_tests/test_redact.py	Adds URL-overlap regression coverage.
ml_tests/test_llm.py	Tests Bedrock client creation and caching.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.