Prompts in LangChain | Generative AI using LangChain | Video 4 | CampusX

LangChain prompts are the control layer that determines what an LLM produces, and the practical way to make that control reliable is to stop asking end users to type full prompts. Instead, the system should collect structured inputs (like paper title, explanation style, and length) and then assemble the final prompt using LangChain’s prompt templates. The transcript also corrects a common misunderstanding about the LLM “temperature” parameter: setting temperature to 0 (or near 0) makes outputs repeatable for the same input, while higher values (around 1.5 in the example) increase variation and creativity.

After a quick recap of earlier LangChain components—especially models—the walkthrough shifts into what “prompts” really mean in practice. Prompts are simply the messages sent to an LLM: text, images, audio, or video can all be used depending on the model. For the current lesson, the focus stays on text-based prompts, which remain the most common approach. Because LLM outputs are highly sensitive to prompt wording, prompt engineering has become its own job category, and LangChain provides tools to design prompts more safely and consistently.

A major distinction follows: static prompts versus dynamic prompts. Static prompting means the user writes the entire instruction (for example, “Summarize this research paper in five lines”). That approach is fragile—misspellings, wrong paper names, or even small wording changes can lead to unpredictable or undesirable results. The transcript argues that real applications need consistency across users, so the better pattern is dynamic prompting: a template is created once, and runtime variables fill in the blanks.

To demonstrate, the lesson builds a small Streamlit “research assistant” app. The UI collects three structured inputs: a paper selection (via dropdown), an explanation style (e.g., code-heavy, math-heavy, simple), and a summary length (short/medium/long). A LangChain PromptTemplate then injects these values into a prewritten template that also enforces requirements like including relevant mathematical equations when present, using intuitive explanations, and ensuring the output matches the requested style and length. The transcript further justifies PromptTemplate over raw f-strings: templates provide built-in validation (catching missing or extra placeholders early) and improve reusability by allowing templates to be saved and loaded from separate files.

Next comes multi-turn chat. A simple console chat bot initially fails at context because it sends only the latest user message, so follow-up questions (like “multiply the bigger number by 10”) can’t be resolved. The fix is to maintain chat history and send the full conversation each time. The transcript then introduces LangChain’s message types—SystemMessage, HumanMessage, and AIMessage—and shows how labeling each message solves ambiguity as the conversation grows.

Finally, the lesson extends the same template philosophy to more advanced prompt construction: ChatPromptTemplate for dynamic sets of messages, and MessagePlaceholder for inserting stored chat history into a prompt at runtime. The result is a blueprint for building LLM applications that remain consistent, context-aware, and maintainable as conversations and requirements scale.