trollknurr / hard-chill-response Goto Github PK

"As a user, I want to interact with the bot in the simplest possible way, using templates / ready-made promts / "simple" promts and get the most complete and accurate result"

Based on research here and here

1. Before user input - hidden prepromt for model role "as an expert in next domain, ..." ("first chains" ref here https://www.promptingguide.ai/ru/techniques/prompt_chaining or like here https://github.com/rodion-m/awesome_ai_for_programmers?tab=readme-ov-file#царь-промпты)
2. Healthy prompt injections - integrate ref here https://github.com/rodion-m/awesome_ai_for_programmers?tab=readme-ov-file#промпт-тюнинг
3. Input format - 1-3 steps with pre-made buttons with already prepared prompts to set up a follow-up model response
4. Input format - additional step with free text input
5. Output format - pre-made buttons for text and table variants of output - maybe part of pnt. 2

Run tests on pull requests

Check code quality withmake lint command

User ask question somehow (telegram for now) and recevie an comprehensive answer, with documents or citations from them

Succesfully found one small inconvenience provided by using pydantic dataclasses.

Not a big news, because JetBrains already aware of this case, bring the proof.

As a workaround you can install Pydantic plugin to Pycharm for now, not a best solution but it brings away annoying yellow underline.

Part of #11

Based on QA with RAG user story we definetely want to put bunch of embeddings somewhere 🧺

LangChain already have interfaces for vast amount of vector stores 💝

• Documentation
• Source code

Hexagonal architecture allows us not to think about a specific tool yet ( OpenSearch / Clickhouse / ChromaDb / etc ).

I will wrap up the implementation of ports for interacting with vector storage firstly, and than we can choose solution according of our restrictions and inner wishes.

"As a user, I want to have the most complete and most correct answers based on the documents used"

1. RAG - using advanced flow for processing text and tables in documents (ref here)
2. RAG - implement RAPTOR as a basic arch of all documents (ref here)
3. RAG - getting reference list from sources and their short description + original file name

In telegram we can ingest many types of documents:

application.add_handler(MessageHandler(filters.Document.Category('application/pdf'), downloader))
application.add_handler(MessageHandler(filters.Document.Category('application/msword'), downloader))
application.add_handler(MessageHandler(filters.Document.Category('application/vnd.openxmlformats-officedocument.wordprocessingml.document'), downloader))
application.add_handler(MessageHandler(filters.Document.Category('application/vnd.ms-powerpoint'), downloader))
application.add_handler(MessageHandler(filters.Document.Category('application/vnd.openxmlformats-officedocument.presentationml.presentation'), downloader))
application.add_handler(MessageHandler(filters.Document.Category('text/plain'), downloader))

Application need input port for documents and implement different adapters for different doc types.

So ports wanted

Definetely want to add some observability to our spaceshiip. And, if LangSmith still in closed beta we should looks for something more open for us.

My suggestions:
https://github.com/Arize-ai/phoenix ( opentelemetry in readmy file and attempt for embedding visualisations )
https://github.com/langfuse/langfuse

@trollknurr
Maybe you know something better or sexier for this kind of intension?

"As a user, I want to be able to upload text documents to the repository and receive responses from the bot according to the information in these text documents"

1. Document upload - shared per client, access via a tg-bot (upload to files - ref https://github.com/sazonovanton/SirChatalot#files )
2. Document upload - single-level storage (without folders), limited set of formats (.txt)
3. Document processing - processing of received documents begins asynchronously to the operation of the tg-bot
4. Document processing - push to the tg-bot with the file name (?) after including it in the storage - ** notifying all users? There will be spam, use timer or separated command in a tg-bot /check_files**
5. Bot answers - automatic updating (?) with the inclusion of document embeddings in the answers - idk, do I need additional squats with a context reboot?

На всякий случай делюсь соской, довольно таки simple, но не слишком cool, потом может принесем с собой https://github.com/Delgan/loguru

#1 (comment)