[Week of 8/21] LangChain Release Notes

New in Retrieval

There was a lot happening in the retrieval space these past two weeks, so we wanted to highlight these explicitly!

MultiVector Retriever: a new retrieval algorithm that enables multiple vector embeddings per document, that can be per-chunk, a summary, hypothetical questions, or more
LangSmith cookbook on evaluating retrieval systems
Our Retrieval Webinar Series continues!
>>Advanced Retrieval with Chroma and Unstructured
>>Production Ingestion with Airbyte and Sweep
>>end-to-end evaluation with Ragas: and a tandem blog post on Evaluating RAG Pipelines
>>In the last webinar series we have Pedro, founder of Tavrn on. We LOVED hearing about how application builders think about and benchmark retrieval systems. If you are building an application and want to join for a future webinar, please reach out at hello@langchain.dev!
New Retrieval Documentation: everything needed for ingestion (load, split, embed, store), plus a collection of retrieval algorithms (self-query, parent document, etc.)
Caching Embeddings: Embeddings can be stored or temporarily cached to avoid needing to recompute them.
Benchmarking Question/Answering Over CSV Data: improving an application that does question answering over CSV data including code and open-sourced eval data, code for gathering feedback, and final agent code here, also on YouTube
Integrating Noah: ChatGPT with Google Drive and Notion data: written in collaboration with the Tavrn team about their new highly personalized, highly context-aware app, built on LangChain.

This week, we’re focusing on cookbooks as part of our effort to help more developers build end-to-end applications.

Use the run_on_dataset helper to benchmark aggregate metrics and check against a threshold
Write individual unit tests to make assertions on every row in a dataset
Make user scores more actionable, with optional comments and corrections
Evaluate your apps via LLM-based preference scoring
Use LangSmith to test your RAG system and make prompt tweaks to improve the chain's performance to improve overall consistency of your LLM applications
if there are other recipes you’d like to see, tell us about them @hello@langchain.dev
Monitoring Charts: Each project now has a monitor tab that allows you to track important metrics over time including trace count, success rate, and latency. We will be adding more metrics very soon!

Added Fallbacks to the LangChain Expression Language (LCEL): a better way to handle LLM API failures in production-ready LLM applications
Caching Embeddings: Embeddings can be stored or temporarily cached to avoid needing to recompute them.
ChatLangChain Improvements: We're benchmarking a bunch of retrieval and agent methods for our "chat langchain" app! Interact with the new beta version here.
Open Source LLM guide: covers open source LLM SOTA (overview fig below) and ways to run them locally (llama.cpp, http://ollama.ai, gpt4all).
MultiVector Retriever: a new retrieval algorithm that enables multiple vector embeddings per document, that can be per-chunk, a summary, hypothetical questions, or more
OpenAI Adapter: we added an easy way to switch out our OpenAI calls for the variety of other models that LangChain supports

Noah (by Tavrn): ChatGPT with your Google Drive and Notion documents
DemoGPT: an open-source project that aspires to keep pushing the boundaries of Large Language Model (LLM) based application development.
GPT Researcher x LangChain: Web research is such a great LLM use-case. To make it easier, we integrated GPT Researcher with our OpenAI adapter, which allows (1) easy usage of other LLM models under the hood, (2) easy logging with LangSmith.
MultiOn bringing the power of agents to the web: using Agents to automate and streamline online interactions.