Why you need a framework at all
A raw OpenAI API call takes three lines. The problem starts when you need call chains, memory, tools, RAG, monitoring, retries, and switching between providers. A framework removes 70% of boilerplate and gives ready-made abstractions. Without it, in production every sneeze turns into custom code.
On reddyx.xyz/feed/ releases of these frameworks are tracked daily, the version release frequency in 2026 is higher than React in 2019.
LangChain: Swiss Army knife with problems
LangChain is the most popular LLM framework, 95k stars on GitHub by May 2026. Strong side, ecosystem: 700+ integrations, ready-made templates for agents, LangGraph for stateful workflow. Weak side, frequent breaking changes and overcomplicated abstractions.
Example: chat with memory in 5 lines
pip install langchain langchain-openai
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage
llm = ChatOpenAI(model="gpt-4o")
response = llm.invoke([HumanMessage(content="Привет")])
print(response.content)
When to take: agent prototype with tools, need 10+ integrations at once, team already knows the API.
LlamaIndex: king of RAG tasks
LlamaIndex is tailored for document indexing and retrieval. If the task is to make a chat over 10000 PDFs, this is the right choice. Built-in parsers for 160+ formats, hybrid search, reranking out of the box. Version 0.12 in early 2026 rewrote the query engine, speed increased by 40%.
pip install llama-index
from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
docs = SimpleDirectoryReader("./data").load_data()
index = VectorStoreIndex.from_documents(docs)
query_engine = index.as_query_engine()
print(query_engine.query("Что в документах про revenue?"))
DSPy: programming instead of prompts
DSPy from Stanford changes the paradigm. Instead of manual prompt engineering you describe the task signature, the framework itself optimizes the prompt via few-shot and bootstrapping. In 2026 this is the fastest-growing framework by GitHub stars (+22k per year).
pip install dspy-ai
import dspy
dspy.settings.configure(lm=dspy.OpenAI(model="gpt-4o-mini"))
classify = dspy.Predict("text -> sentiment")
result = classify(text="Сервис отвратительный, верните деньги")
print(result.sentiment)
When to take: need stable accuracy on a specific task, have labeled data for optimization, reproducibility matters.
Haystack: production without pain
Haystack from deepset focuses on reliability. Pipeline as a graph, strict data types, native integration with Elasticsearch and OpenSearch. Version 2.x released in 2024, by 2026 it's a mature product used in banks and insurance.
pip install haystack-ai
from haystack import Pipeline
from haystack.components.generators import OpenAIGenerator
pipe = Pipeline()
pipe.add_component("llm", OpenAIGenerator(model="gpt-4o"))
result = pipe.run({"llm": {"prompt": "Объясни RAG в одном предложении"}})
print(result["llm"]["replies"][0])
Semantic Kernel: choice for .NET
Semantic Kernel from Microsoft, the only serious framework with equal support for Python, C# and Java. If the team is on .NET, there are no alternatives. Plugins, planners, native integration with Azure OpenAI. Minus, documentation sometimes lags behind code by a month or two.
Comparison table
| Framework | GitHub Stars | Best for | Weak spot |
|---|---|---|---|
| LangChain | 95k | Agents, prototypes | Breaking changes |
| LlamaIndex | 38k | RAG, document search | Fewer agent abstractions |
| DSPy | 22k | Prompt optimization | Learning curve |
| Haystack | 17k | Production RAG | Smaller community |
| Semantic Kernel | 22k | .NET stack | Documentation lags |
What to choose in 2026
If building an MVP over a weekend, LangChain. Chat over documents, LlamaIndex. Serious task with quality metrics, DSPy. Production with SLA, Haystack. Corporate on .NET, Semantic Kernel. Combining is also normal, many use LlamaIndex for retrieval inside a LangChain agent.
New releases of all these frameworks appear in the catalog reddyx.xyz/feed/ automatically.
Frequently Asked Questions
Which LLM framework is the most popular in 2026?
LangChain with 95k stars on GitHub, but DSPy is growing the fastest, +22k stars in the last year.
Can multiple frameworks be used simultaneously?
Yes, it's a common practice: LlamaIndex for retrieval plus LangChain for agent orchestration.
Which framework is better for RAG?
LlamaIndex is tailored for RAG, has 160+ document parsers and hybrid search out of the box.