I was initially thrilled to hear about the new features announced at AWS re:Invent, particularly regarding Amazon Q and expanded LLM support in Bedrock. However, my experience over the past two days has been disappointing. Here are some specifics:
Amazon Q with Native Retriever and S3: The synchronisation feature that crawls and indexes data is disappointingly slow. For instance, crawling a mere 5 KB text file took an unexpectedly long time, and indexing continued for over two hours before I stopped it.
Kendra Retriever for Indexing a Confluence Page: Here, the crawling of a single Confluence page using a regex filter was somewhat faster, completing in about 20 minutes. However, the indexing is still ongoing, and I plan to let it finish to see the total duration. Interestingly, despite the indexing process being incomplete, I was able to query and retrieve results from the indexed content via the Kendra console.
Issues with Kendra Retriever on Amazon Q: I also attempted to use the Kendra retriever with Amazon Q, but it didn’t work. This might be due to incomplete indexing, or perhaps there's something I'm missing in the setup.
For context, six months ago, I developed a similar interface and backend system. It was designed for uploading various file types (including PDFs requiring extensive OCR, text files, docs, etc.), scanning Confluence spaces and pages, and crawling web pages. System that I built using open source libraries like Langchain, Unstrcutured.io in a serverless fashion, indexes content in Qdrant vector DB. This setup powers our internal chatbot using #RAG, assisting our sales and support teams in quickly finding answers. The efficiency and speed of this in house built system stand in stark contrast to my current experience with Amazon Q and its associated tools.
I think AWS can make it more simpler for anyone to use, build and release their own custom chatbots with their own specific business domain knowledge. Will keep myself tuned.
#AWS #reinvent2023 #qdrant #RAG #langchain #genai