Memory support for chains with multi inputs Previously, memory only worked if there was single input This meant it couldn't work in cases like question/answering, where you passed in multiple inputs (documents and a question) Now it can: https://
langchain.readthedocs.io/en/latest/exam
ples/memory/adding_memory_chain_multiple_inputs.html
…
LLMS
-
Memory Support Now Works with Multiple Chain Inputs
By
–
-
LangChain v0.0.51 Release: Enhanced Tools, Memory, and Documentation
By
–
v0.0.51 Unify return types for intermediate steps – h/t @johnjnay Add sports & knowledge graph results to SerpAPI wrapper – @brunotorious Document improvements: @eltociear @samcwl Making tools editable – h/t @brunotorious Memory support for chains with multi inputs
-
OT Too Low Level: Codex Better for Code Refactoring
By
–
OT is way too low level man. too much structure to have to rebuild when codex can already do entire coherent code blocks of refactoring
-
Embeddings as Versioned Documentation for Machine-Readable Code
By
–
my take: Docs for humans, Embeddings for machines Write samples -> get embeddings -> ship with each API/SDK version composable give users 99% accurate codegen can't version foundation models, but can version embeddings If you liked @types
, you'll *LOVE* @embeddings! -
Beyond Copilot: The Next Evolution in AI Developer Experience
By
–
The Big Question in AI-driven developer experience: What's after Copilot? Everyone is convinced this is "just" the beginning but can't articulate what the next killer form factor will be. Increasingly convinced next step change won't be from optimal stuffing context windows
-
Customizable AI Response Prefix in Conversational Memory
By
–
More customizability for conversational memory A small change, but allows for specification of the prefix to use for for the AI response Was previously hardcoded to "AI", but can now be set to whatever Idea came from the discord! https://
langchain.readthedocs.io/en/latest/exam
ples/memory/conversational_customization.html
… -
New Memory Type Combining Summary and Context Window
By
–
New memory type Previously, we had one memory type that created a summary, another that kept a window of recent lines of conversation. jaloo555 added a new type of memory which does both! It decides the window based on token length. https://langchain.readthedocs.io/en/latest/examples/memory/conversational_memory.html#ConversationSummaryBufferMemory
-
Map-Reduce and Refine Chains Now Return Intermediate Steps
By
–
Return intermediate steps for map-reduce & refine chains @johnjnay had the great idea to allow for returning of intermediate steps of these chains to the user This makes it easy to inspect things like Q&A over each document, before combining https://
langchain.readthedocs.io/en/latest/exam
ples/data_augmented_generation/question_answering.html#The-map_reduce-Chain
… -
Optimizing minGPT: Performance improvements from 495ms to 102ms
By
–
having fun optimizing minGPT today
– base: 495ms
– zero_grad(set_to_none=True): 492
– torch.jit.script gelu: 463
– OMP_PROC_BIND=CLOSE: 453
– torch.backends.cuda.matmul.allow_tf32: 143
– torch.autocast(torch.bfloat16): 121
– FlashAttention: 102
now: more fused kernels more better -
Splitting minGPT into educational and efficient versions
By
–
Context I realized I have to split up minGPT because I can't properly simultaneously satisfy both 1) educational and 2) efficient in one repo. So I'm separately writing 1) the maximally educational minGPT (+video etc.) and 2) a more efficient (still ~clean) version that has teeth