Thanks Tammy! I created a DSPy pipeline for question answering:
- 1.
- 2.
Determine type of ‘artifact’ to construct:
- a.
If related to time, focus on building a timeline of events
- b.
If not, focus on gathering facts to build an outline
- 3.
Chain of Thought: Create search query ->
- 4.
Retriever 1: Get relevant articles from ColBERTv2 as context ->
- 5.
Chain of Thought: Create additional search query to fill in gaps ->
- 6.
Retriever 2: Get relevant articles from You.com and add to context ->
- 7.
Construct the artifact from the combined context ->
- 8.
Answer the question from the created artifact
There was a bit of complexity here, so it was nice being able to explore the chain in Phoenix. Much more user friendly than lm.inspect !