Oct 2022
LangChain ships: "everything connected" LangChain
Harrison Chase's original bet: make it trivial to chain LLM calls with any provider, any tool. Rapid integrations means rapid dep growth. The philosophy was breadth-first — connect everything, worry about weight later.
Dec 2022
LlamaIndex: "all-in-one data platform" bet LlamaIndex
Jerry Liu's founding thesis: retrieval-augmented generation needs a complete data stack — loaders, splitters, indices, storage, querying. Bundling SQLAlchemy, NLTK, Jinja2, and Pillow follows logically from that thesis.
2023
Pydantic v2 migration fractures the ecosystem Both
Pydantic's v1→v2 breaking change exposed the hidden cost of shared deps. Both LangChain and LlamaIndex needed to update simultaneously. Projects using both saw broken installs for months. The pin wars began.
Jan 2024
LangChain 0.1: ecosystem consolidation LangChain
LangGraph and LangSmith become core deps — bundled whether you want them or not. Dep count stabilizes around 32 unique packages. The bet shifts from "integrate everything" to "own the full stack."
Mar 2024
SynapseKit: the 2-dep constraint as architecture SynapseKit
Designed with a hard rule: ship only numpy (vector math) and rank-bm25 (BM25 retrieval) at the core. All provider deps are optional extras. The founding bet: the right abstraction doesn't require pulling in the world.
2025–2026
Dependency footprint becomes a selection criterion Ecosystem
As teams use LLM frameworks in existing codebases — not just greenfield projects — version conflicts, security scan surface, and install time become real selection criteria. Footprint matters.