COURSE · FND · B011
LangChain for Local AI
Learn langchain for local ai through RunLocalAI's practical lens: langchain, chains, rag and prompts, hardware fit, runtime settings, verification habits and local-vs-cloud tradeoffs.
PREREQUISITES
- B002
- B003
Course B011: LangChain for Local AI
Why this course exists
LangChain provides the compositional layer that turns raw LLM API calls into stateful, multi-step pipelines. This course exists because local AI deployments—using Ollama, llama.cpp, or similar runtimes—require you to wire these components yourself, without the managed infrastructure that cloud providers abstract away. If you have completed B002 (Local LLM fundamentals) and B003 (CLI and API usage for local models), you have working models. Now you need to compose them into chains. This course bridges that gap. Skip to Chapter 3 if you already understand LangChain's value proposition.
What you will know after
- Construct prompt templates that accept runtime variables instead of hardcoded strings
- Connect multiple LLM calls in sequence to process multi-step tasks
- Route user queries to domain-specific chains without if/else branching
- Parse raw model output into structured Python objects (dicts, Pydantic models, enums)
- Maintain conversation history across multiple turns using built-in memory components
- Use LangChain's Ollama integration to run all of the above against local models
CHAPTERS
- 01What is LangChain?LangChain's Chain abstraction standardizes how you compose LLM calls with prompts, memory, and output parsers into reusable, swappable pipelines.20 min
- 02Ollama LLM IntegrationThe `langchain-ollama` integration connects LangChain to a local Ollama HTTP server; verify Ollama is running and the model is loaded before initializing `ChatOllama`.25 min
- 03Prompt Templates`PromptTemplate` separates prompt structure from prompt data, allowing you to reuse template strings with different runtime values across multiple chain invocations.20 min
- 04ChatPromptTemplate`ChatPromptTemplate` builds role-labeled message sequences that chat models consume natively, contrasting with `PromptTemplate`'s single-string approach.25 min
- 05LLMChain Basics`LLMChain` binds a prompt template to an LLM into a single callable, executing the full render → invoke → parse pipeline in one `.invoke()` call.20 min
- 06Sequential Chains`SequentialChain` wires multiple LLM calls in series with explicit input/output variable mapping, enabling multi-step pipelines like classify-then-summarize or extract-then-expand.20 min
- 07Router ChainsRouter chains replace branching logic with a model-driven dispatcher, pointing queries to domain-specific chains based on the LLM's classification.20 min
- 08LangChain Output ParsersOutput parsers decouple the LLM's freeform text output from your application's type system, wrapping extraction and retry logic so your code receives typed Python objects.25 min
- 09Memory: ConversationBuffer`ConversationBufferMemory` persists message history across turns by storing and replaying it with each chain invocation, enabling stateful multi-turn conversations.25 min
- 10Memory: SummarySummary memory trades perfect recall for bounded memory usage—essential for long-running conversations on resource-constrained local deployments.15 min
- 11Memory: Vector StoreVector store memory retrieves semantically relevant conversation history on demand, enabling the model to recall specific past interactions without bloating context windows.15 min
- 12Document LoadersDocument loaders normalize disparate file formats into LangChain's `Document` object, providing a unified interface for downstream processing.20 min
- 13Text SplittersText splitters chunk documents for embedding while `chunk_overlap` maintains cross-chunk context—setting chunk_size to 20-30% of your embedding model's context yields the best retrieval results.20 min
- 14Simple RAG PipelineRAG pipelines separate knowledge storage (vector database) from knowledge application (LLM generation), enabling accurate responses grounded in your documents.20 min
- 15RetrievalQA Chain`stuff` works for <4 documents, `map_reduce` for large batches, and `refine` when document order matters—choose based on document count and whether order affects meaning.20 min
- 16Streaming with LangChainStreaming requires explicit `stream()` calls—`invoke()` always waits for complete generation even with `stream=True` in the LLM config.20 min
- 17LangChain CallbacksCallbacks intercept chain execution events without modifying chain logic—ideal for production observability and debugging.20 min
- 18LangChain EvaluationWithout evaluation metrics, you cannot distinguish improvements from regressions—build evaluators for every production chain.20 min