Add Streaming Response for Jupyter/REPL
In Jupyter, showing partial responses is very UX-friendly. Use LangChain's streaming interface to update a cell incrementally.
Haris Jabbar 10 months ago
Add Streaming Response for Jupyter/REPL
In Jupyter, showing partial responses is very UX-friendly. Use LangChain's streaming interface to update a cell incrementally.
Haris Jabbar 10 months ago
Add Memory / Conversation History
Right now each question is stateless. Add an optional conversation memory (e.g., past 3 Q&A pairs) in AgentState["messages"], and pass them to the LLM for better coherence.
Haris Jabbar 10 months ago
Add Memory / Conversation History
Right now each question is stateless. Add an optional conversation memory (e.g., past 3 Q&A pairs) in AgentState["messages"], and pass them to the LLM for better coherence.
Haris Jabbar 10 months ago
Better prompts
Currently the prompts are quite vanilla. There is significant room for improvement for better prompt templates.
Haris Jabbar 10 months ago
Better prompts
Currently the prompts are quite vanilla. There is significant room for improvement for better prompt templates.
Haris Jabbar 10 months ago
Async & Parallel Context Gathering
Currently everything is sequential. Instead: Gather traceback, module info, and context summary concurrently. Use async def and asyncio.gather() for performance (especially in Jupyter).
Haris Jabbar 10 months ago
Async & Parallel Context Gathering
Currently everything is sequential. Instead: Gather traceback, module info, and context summary concurrently. Use async def and asyncio.gather() for performance (especially in Jupyter).
Haris Jabbar 10 months ago
Improved context for large modules
Currently, we add docs for all submodules as context. This can quickly get expensive or even run out of context window. We need to do smart context engineering by using RAG/Summarization/Selective filtering or a combination of these.
Haris Jabbar 10 months ago
Improved context for large modules
Currently, we add docs for all submodules as context. This can quickly get expensive or even run out of context window. We need to do smart context engineering by using RAG/Summarization/Selective filtering or a combination of these.
Haris Jabbar 10 months ago