ToDoList.md

Project Todo List

Tech Stack used

• Autogen
• Models :
  • LLM
    • GPT-3.5-turbo (OpenAI)
    • Claude-2.0 (Anthropic)
    • Gemma-2b-it (Google)
  • Embedding
    • mxbai-embed-large-v1 (Mixedbread AI)
  • CrossEmbedding
    • mxbai-rerank-base-v1 (Mixedbread AI)
  • Safety
    • Toxicity
      • bias-detection-model (d4data)
    • Bias
      • bias-detection-model (d4data)

Project : Coder Assistants

• Code Assistant agent templates

  • Python Dev Team

    • Co-ordinating Manager
      • Technical Manager: Identifies which agent is required for the successive tasks.
    • Developer Team
      • Project Planner: Breaks down the initial problem into simpler chunks to be done by the code writer
      • Code Writer: Writes codes
      • QA developer: Write code tests to be done
    • Support Tools
      • Code Repo: Identifies already existing code repository that can be used for development.
      • Code Testing support: Runs and checks if the codes work as intended.

  • SQL Dev Team

    • Co-ordinating Manager
      • Technical Manager: Identifies which agent is required for the successive tasks.
    • Developer Team
      • Lead Analyst: Breaks down the complex issues in smaller queries that can be used by the Senior analyst and Junior Analyst.
      • Senior Analyst: Write queries for complex tasks
      • Junior Analyst: Write queries for simpler tasks
    • Support Tools
      • Database Admin: Helps the Analyst team to write queries by providing context about the tables and columns that can be used for building queries.
      • Query Testing support: Runs and checks if the queries work as intended.

• Code Migration

  • SAS to PySpark (TBD)

    • Create a SQL Code and Assets repository (borrows Database Admin from SQL Dev team)

      • Tables and related metadata
        • Tables and their description
        • Tables Columns and their description [comment]: # (This resembles closely to a data dictionary.)
      • Relationships between tables
        • Join conditions from existing query repo
        • Keys mapping (FK,CK,PK etc)
      • Pre-existing queries as baseline
        • Queries for specific request as example (especially if request is closely related or similar)

    • Create a PySpark Code repository (TBD)

  • SAS to SQL (TBD)

  • Python(Pandas) to PySpark (TBD)

Learning for future updates

• Using LiteLLM for supporting multiple API endpoints (https://docs.litellm.ai/)
• Check how Langfuse works (https://langfuse.com/)
  • For monitoring of the API use
• Control vectors and Representation Engineering
  • Control vectors (https://www.reddit.com/r/LocalLLaMA/comments/1atqj7f/control_vectors_add_a_meaningful_bias_in_each/)
  • Representation Engineering (https://vgel.me/posts/representation-engineering/)
• Running LLM locally
  • Ollama (https://github.com/ollama/ollama)
  • Llama.cpp (https://github.com/abetlen/llama-cpp-python)
• Simplify tool usage for agents
  • Semantic Router (https://github.com/aurelio-labs/semantic-router)
    • Could also be used for testing topics that needs to be guarded against.
      • Politics and/or Political Figures
      • Sensitive information (Creating explosives, biohazards etc.)
      • PII
      • HIPPA (Health, Insurance related)
    • Identify which agent / tool can be used for the next steps (building own agent framework?) =======