NA
DOCUMENT

Natural Language Processing with Large Language Models Cheat Sheet

A large-language-model NLP reference for text cleaning, tokenization, prompt design, embeddings, retrieval, generation evaluation and privacy.

Version 2026-08-03通用Public reference material; verify the included notice and project terms before redistribution

What this reference covers

This NLP reference breaks text work into data preparation, model calls, prompting, embeddings, retrieval, structured generation and evaluation. It is designed to help a developer make a complete input-to-output workflow rather than solve every task with a longer prompt.

Define the task and failure cost

Distinguish classification, extraction, summarization, question answering and generation. State input fields, output schema, refusal conditions and acceptable errors. Preserve source, time, permission and paragraph identifiers when long documents are cleaned or split.

Choose prompting or retrieval deliberately

Simple tasks may use a structured prompt. Knowledge-heavy tasks may add embeddings and retrieval, but retrieved text must actually support the answer. Evaluate recall, citation completeness, factual consistency, stability, cost and sensitive-data exposure.

Maintenance note

Keep minimal logs and access policies for inputs, embeddings, indexes and outputs. Fluent generated text is not a fact database. Content review date: 2026-08-23.

SAVE TO CLOUD

Save to your cloud drive

Save the complete collection first so files remain together and are easier to access across devices.

Links checked 2026-08-06
Save first, access when you need itOn desktop, scan with the matching cloud-drive app. On mobile, tap the save button.
GUIDE

LLM NLP study guide

Define the task and error cost, prepare de-identified samples, choose prompting or retrieval based on evidence and preserve failed cases for evaluation.

Before you start

  • Prepare a PDF reader and document text sources, language, length and sensitive fields.
  • Prepare representative de-identified samples and a human review rubric.
  • Decide the output schema, refusal conditions and citation expectations.
02

Quick start

  1. 01

    Define task and output

    Separate classification, extraction, summary, question answering and generation, and write input fields, output schema and acceptable errors.

  2. 02

    Clean and segment text

    Normalize encoding, remove unnecessary repetition, split long text by meaning and preserve source and paragraph identifiers.

  3. 03

    Choose prompt or retrieval

    Use structured prompting for simple tasks and add embeddings and retrieval for knowledge-heavy tasks, checking whether retrieved text supports the answer.

  4. 04

    Evaluate failures

    Measure correctness, completeness, citations, stability, cost and sensitive-data exposure, and preserve failed samples with configuration versions.

Usage tips

  • Treat generated text as a hypothesis and return important claims to verifiable source material.
  • Keep title, date and permission metadata with retrieved passages to avoid context-free answers.
  • Set minimal retention and access rules separately for user input, logs, embeddings, indexes and outputs.
Troubleshooting and uninstall

What if an answer sounds reasonable but is factually wrong?

Add representative evaluation cases and refusal conditions, inspect retrieval and prompt constraints and route important claims through a human or rule check.

Why are long-document retrieval results irrelevant?

Check chunk size, overlap, metadata, query rewriting and retrieval count with labelled examples before adjusting generation instructions.

FAQ

Frequently asked questions

What tasks does the LLM NLP sheet support?

It supports workflow lookup for classification, extraction, summarization, question answering, embeddings and retrieval-augmented generation.

When should retrieval augmentation be considered?

Consider it when answers rely on changing, large or permission-controlled knowledge, and evaluate relevance and completeness of retrieved material.

How should generated results be evaluated?

Combine task metrics, human review, factual consistency, refusal quality, stability, cost and privacy checks instead of language fluency alone.

What should be checked before uploading text?

Identify personal data, internal documents, credentials and copyrighted material, then de-identify, limit retention and confirm processing boundaries.