Skip to content

Latest commit

 

History

History
64 lines (56 loc) · 3.51 KB

TODO.md

File metadata and controls

64 lines (56 loc) · 3.51 KB
Raw

TODO

Done-ish

  • Update all namespaces
  • Sketch out Knowledge class, it is essentially an abstraction around splitting, embedding and indexing a piece of
  • Separate classes per loader, registry pattern, allows overriding a loader (ex Pdf loader can use Textract OCR instead of PDF Parser etc)
  • Figure out what vectorstore interface should look like
    • We need these features:
      • Create an index of a certain dimension
      • Delete an index??
      • Insert a single vector into a "collection" and specify an id (with metadata)
      • Upsert a single vector into a "collection" and specify an id (with metadata)
      • Insert multiple vectors into a "collection" and specify an id (with metadata)
      • Upsert multiple vectors into a "collection" and specify an id (with metadata)
      • Need class to represent a single result from vectorstore (aka Result), contains:
        • score (float)
        • id (string)
        • meta (array)
        • vector (EmbeddingVector)
  • InMemory (array) vectorstore driver for testing and "throwaway use"
  • expand LLM interface, add some other common llms to dogfood a suitable interface on.
  • Pinecone vectorstore driver ( index = collection, meta = meta) + tests
  • "File" vectorstore driver (Stores everything as JSON file, very stupid and naive solution, for local dev and testing)
  • Command to generate a Tool class (php artisan mindwave:tool)
  • Sketch out Brain class that is an abstraction around vectorstores that operate on Documents
  • Write some simple tests that shows that an agent can answer a simple question based on the content of a text file.

Vectorstore Drivers

TODO

  • Document loader for text/csv
  • Document loader for text/xml
  • Document loader for text/calendar (.ical files)
  • Document Loader for Excel file
  • Tool to search your Gmail (provide access token or service account in constructor, make example in demo app on how to do that in practice)
  • Laravel Scout-ish feature where you can modify your Model to make it indexable, and toPrompt() method to describe text, with converters for other filetypes (ex: pdf -> text))
  • Port the different "agent types" from LangChain, but simplify them.
  • Chain: QA With sources ( inspiration: https://github.com/hwchase17/langchain/blob/master/langchain/chains/qa_with_sources/stuff_prompt.py)
  • Chain: Related to above ^: QA With Source and Eloquent Model: Combine embedding search, model references and retrieve the llm answer along with a model instance as result
  • Add eloquent backend ChatHistoryMemory (migration, model, config file (use mindwave.php) -> chat_history (id, role, message, meta, created_at, updated_at)) the model in natural language or as a CSV
  • Callbacks implemented as Events (base the names on these )

Notes & thoughts

  • Is a Brain a tool, does it make sense to implement it as a tool? (langchain has the concept of retrievers, which is similar to brains)