-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
In Langroid (the agent-oriented LLM framework from ex-CMU/UW-Madison researchers), we call it Relevance Extraction — given a passage and a query, use the LLM to extract only the portions relevant to the query. In a RAG pipeline where you optimistically retrieve top k chunks (to improve recall), the chunks could be large and hence contain irrelevant/distracting text. We concurrently do relevance extraction from these k chunks: https://github.com/langroid/langroid/blob/main/langroid/agen...
One thing often missed in this is the un-necessary cost (latency and token-cost) of parroting out verbatim text from context. In Langroid we use a numbering trick to mitigate this: pre-annotate the passage sentences with numbers, and ask the LLM to simply specify the relevant sentence-numbers. We have an elegant implementation of this in our RelevanceExtractorAgent using tools/function-calling.
Here's a post I wrote about comparing Langroid's method with LangChain's naive equivalent of relevance extraction called `LLMChainExtractor.compress` , and no surprise Langroid's methos is far faster and cheaper: