Prompt Engineering for Chatgpt (2)

 2. Introducing New Information to the Large Language Model 

3. How to deal with Prompt Size Limitations ?

✅ 1. Query Only What’s Relevant (Selective Retrieval)

Idea: Instead of loading entire documents, pull only the parts that matter to the task.

Example:
You're working with a collection of meeting transcripts and want insights on a decision about budget planning.

Instead of this (too long):

“Here are 20 pages of meeting notes…”

Do this:

Use a search function or embedding model to extract only the paragraphs mentioning “budget,” “finance,” or “cost estimate.”

Prompt to LLM:

“Here are excerpts from recent meetings related to budget planning. Please summarize the decisions made and any unresolved issues.”

---

✅ 2. Filter Out Extraneous Information

Idea: Manually or programmatically remove boilerplate, irrelevant sections, or repeated text before prompting the model.

Example:
You have product reviews, but many just repeat phrases like “great product” or include unrelated shipping complaints.

Instead of this:

All reviews, including filler or duplicate content.

Do this:

Pre-filter to keep only reviews that mention specific features (e.g., battery life, camera quality).

Prompt to LLM:

“Summarize customer feedback about the battery life and camera quality based on these filtered reviews.”

---

✅ 3. Summarize or Compress in Chunks (Progressive Summarization)

Idea: If you have a long document, break it into parts, summarize each, then feed the summaries to the model for final reasoning.

Example:
You have a 50-page clinical study report you want to analyze.

Step 1 Prompt:

“Summarize the key findings and methodology of Section 1 of this study.”

(Repeat for all major sections.)

Step 2 Prompt:

“Based on the summaries of all sections, provide a final overview highlighting safety concerns and study limitations.”

Bonus tip: You can even ask the model to preserve specific elements during summarization, like “retain statistical outcomes and participant demographics.”

---

🔁 Summary of the Three Approaches

Approach	What You Do	Why It Works
1. Selective Querying	Pull only relevant parts	Avoids feeding unnecessary info
2. Filtering	Remove noise or irrelevant data	Keeps prompt lean and focused
3. Summarizing in Chunks	Compress data step-by-step	Enables reasoning across large documents

These strategies can extend the power of your prompts and help you work around model limits—especially when dealing with long documents, noisy data, or multi-step tasks.