updated readme

Author: Snaiel

README.md

@@ -1,27 +1,32 @@
 # OpenAI-Milvus-QA-Over-Docs
 
-![screenshot of prgram](images/screenshot.png)
+![screenshot of program](images/screenshot.png)
 
-Uses [Milvus](https://milvus.io/) as a document store and [OpenAI's](https://platform.openai.com/docs/models/gpt-3-5) chat API for a simple app that allows the user ask question based on given sources.
+Uses [Milvus](https://milvus.io/) as a document store and [OpenAI's](https://platform.openai.com/docs/models/gpt-3-5) chat API for a simple app that allows the user ask question based on given sources. Stores every question asked and answer generated in an SQLite relational database which provides additional functionality and analysis.
 
 - Supports CSV, PDF files
 - Supports web pages
-- Provides the maximum possible context to the AI model
+- Shows sources in response
+- Aims to provide as much context to the AI model
+- Stores all questions and answers
 - No chat memory/history
 
 ## How it works
 
 ![diagram of architecture](images/diagram.png)
 
-1. A Milvus instance is run
-2. Files and websites are ingested through [Langchain's](https://github.com/hwchase17/langchain) document loaders and text splitter
-3. Documents embedded by OpenAI embeddings and added to Milvus collection through `langchain`
-4. Only data ingestion done through Langchain, rest uses `pymilvus` and `openai`
-5. A user inputs a query into the chat interface, and gets embedded by OpenAI embeddings (okay embeddings still done through `langchain`)
-6. Similarity search is done with the embedded query and the top 20 most similar documents are returned
-7. From the top 20, as much context/text is retrieved until the token limit is reached. 4096 for OpenAI gpt-3.5 (maximum set to 4000)
-8. Instructions for the model, the context, and the original question is given to the OpenAI chat model
-9. Response is returned and displayed in a chat interface
+1. A Milvus and SQLite instance is run
+3. Files and websites are ingested through [Langchain's](https://github.com/hwchase17/langchain) document loaders and text splitter
+4. Documents embedded by OpenAI embeddings and added to Milvus collection through `langchain`. References of these documents are put in an SQLite table.
+5. Only data ingestion done through Langchain, rest uses `pymilvus` and `openai`
+6. A user inputs a query into the chat interface, and gets embedded by OpenAI embeddings (okay embeddings still done through `langchain`)
+7. Similarity search is done with the embedded query on previous questions
+8. Using distance, it deems whether the question is similar or identical to a previous question.
+9. If so, retrieve the previous response
+10. If there are no relevant questions or users specifies to generate a new answer, do a similarity search on the sources and retrieve the top 20 most similar documents
+11. From the top 20, as much context/text is retrieved until the token limit is reached. 4096 for OpenAI gpt-3.5 (maximum set to 3700 to leave room for response tokens)
+12. Instructions, the context, and the original question is given to the OpenAI chat model
+13. Response is returned and displayed in a chat interface
 
 ## How to run