Prompt & Context Engineering Interview Questions Hindi mein

Prompt और Context Engineering कोर्स के 25 सवाल, हर सवाल के सही जवाब और explanation के साथ। जवाब देखने से पहले ख़ुद सोचें।

Prompts से Context तक

1. Context engineering prompt के शब्दों से आगे क्या जोड़ती है?

  1. मॉडल के weights प्रशिक्षित करना
  2. अच्छा font चुनना
  3. तय करना कि window में कौन-सी सामग्री, किस क्रम, संरचना और budget में जाए
  4. Temperature बढ़ाना
जवाब देखें

जवाब: तय करना कि window में कौन-सी सामग्री, किस क्रम, संरचना और budget में जाए

जोड़ा गया इनपुट तय करता है कि उस call पर मॉडल क्या जान सकता है।

पाठ: Context Engineering का अर्थ

2. Context window में क्या गिना जाता है?

  1. इनपुट tokens और उत्पन्न आउटपुट tokens
  2. सिर्फ़ उपयोगकर्ता का प्रश्न
  3. सिर्फ़ system prompt
  4. सिर्फ़ images
जवाब देखें

जवाब: इनपुट tokens और उत्पन्न आउटपुट tokens

उत्तर के लिए भी जगह रखें।

पाठ: Tokens और Context Window

3. किस खंड को अविश्वसनीय मानना चाहिए?

  1. आपके निश्चित उदाहरण
  2. आपके समीक्षित system नियम
  3. आपकी tool परिभाषाएँ
  4. Retrieved दस्तावेज़ और tool परिणाम
जवाब देखें

जवाब: Retrieved दस्तावेज़ और tool परिणाम

आपके code के बाहर से आई सामग्री में कुछ भी हो सकता है, निर्देश समेत।

पाठ: Model Request की संरचना

Window का Budget

4. आम तौर पर क्या कभी नहीं काटना चाहिए?

  1. कम score वाले दस्तावेज़
  2. System नियम और वर्तमान प्रश्न
  3. बहुत पुराने history turns
  4. अतिरिक्त उदाहरण
जवाब देखें

जवाब: System नियम और वर्तमान प्रश्न

सबसे कम मूल्यवान सामग्री पहले काटें।

पाठ: Token Budget का आवंटन

5. बहुत लंबा context उत्तरों को नुक़सान क्यों पहुँचा सकता है?

  1. मॉडल हमेशा लंबे इनपुट मना करते हैं
  2. प्रासंगिक तथ्य भटकाने वाली चीज़ों में दब जाते हैं, और लागत व latency बढ़ती है
  3. लंबे इनपुट मॉडल weights बदलते हैं
  4. Tokens मुफ़्त हो जाते हैं
जवाब देखें

जवाब: प्रासंगिक तथ्य भटकाने वाली चीज़ों में दब जाते हैं, और लागत व latency बढ़ती है

सबसे छोटा अत्यधिक प्रासंगिक set भेजें।

पाठ: लंबा Context हमेशा बेहतर क्यों नहीं

6. उत्तर लंबाई सीमा पर रुका। क्या सच है?

  1. इसका अर्थ है इनपुट ख़ाली था
  2. यह हमेशा पूर्ण है
  3. इसका अर्थ है मॉडल ने मना किया
  4. यह अधूरा हो सकता है और सँभालना चाहिए, अंतिम नहीं मानना
जवाब देखें

जवाब: यह अधूरा हो सकता है और सँभालना चाहिए, अंतिम नहीं मानना

आउटपुट पर भरोसे से पहले stop reason जाँचें।

पाठ: आउटपुट आरक्षित और नियंत्रित करना

Selecting What Goes In

7. Why use a minimum relevance score when packing?

  1. Weak matches can mislead the model, so leaving them out is often better
  2. It makes every chunk longer
  3. It increases the context window
  4. It is required by JSON
जवाब देखें

जवाब: Weak matches can mislead the model, so leaving them out is often better

Only include content that genuinely helps.

पाठ: Packing Retrieved Chunks Into a Budget

8. What is a cost of sending duplicate chunks?

  1. They reduce latency
  2. They improve citations automatically
  3. They waste budget and can over-weight one source
  4. They shrink the model
जवाब देखें

जवाब: They waste budget and can over-weight one source

De-duplicate before packing.

पाठ: Removing Near-Duplicates

9. What is an advantage of extractive compression?

  1. It needs a second large model
  2. It always produces the shortest text
  3. Original wording is preserved, so it stays exact and citable
  4. It removes the need for retrieval
जवाब देखें

जवाब: Original wording is preserved, so it stays exact and citable

Extractive keeps exact sentences; abstractive rewrites them.

पाठ: Compressing Content

10. What is dynamic few-shot selection?

  1. Choosing the examples most similar to each request from a larger pool
  2. Using no examples ever
  3. Fine-tuning on every request
  4. Sending the whole pool every time
जवाब देखें

जवाब: Choosing the examples most similar to each request from a larger pool

Relevant examples, shorter prompts.

पाठ: Selecting Few-Shot Examples per Request

Ordering and Structure

11. Where is a sensible default place for the user question in a long context?

  1. Repeated in every document
  2. Hidden in the middle of the documents
  3. Before the system rules
  4. At the end, after the documents
जवाब देखें

जवाब: At the end, after the documents

Question last keeps it close to where the model starts answering.

पाठ: Ordering Content in a Long Context

12. Why escape content placed inside tags?

  1. To make it shorter
  2. So content cannot close the tag and inject its own section
  3. To translate it
  4. To make the model faster
जवाब देखें

जवाब: So content cannot close the tag and inject its own section

Escaping keeps data inside its boundary, though it is not a complete defence.

पाठ: Delimiters and Templates

13. Why add a short reminder near the question in long contexts?

  1. It replaces the system prompt
  2. It helps the model apply key constraints after reading a lot of content
  3. It doubles the context window
  4. It removes the need for evaluation
जवाब देखें

जवाब: It helps the model apply key constraints after reading a lot of content

A brief restatement of the critical rules is cheap and often helps.

पाठ: Instruction Placement and Repetition

History, Memory and Tool Output

14. Why not send the full conversation forever?

  1. It makes answers shorter
  2. Models cannot read history
  3. History is illegal to store
  4. It eventually overflows the window and increases cost every turn
जवाब देखें

जवाब: It eventually overflows the window and increases cost every turn

Window plus summary keeps context bounded.

पाठ: Managing Conversation History

15. What should happen to long-term memories over time?

  1. They should be updatable and deletable, with privacy rules applied
  2. They should never change
  3. They should all be sent on every call
  4. They should include passwords for convenience
जवाब देखें

जवाब: They should be updatable and deletable, with privacy rules applied

Memory is user data with a lifecycle.

पाठ: Long-Term Memory: Store and Retrieve

16. Why keep the tail when truncating a long log?

  1. Models only read the tail
  2. The tail is always shorter
  3. Errors and final status often appear at the end
  4. The head is never useful
जवाब देखें

जवाब: Errors and final status often appear at the end

Head-and-tail keeps context and outcome.

पाठ: Keeping Tool Output Under Control

Caching, Output Contracts and Versioning

17. Where should a per-request timestamp go to keep caching effective?

  1. Inside every example
  2. At the very top of the prompt
  3. After the stable instructions and examples
  4. In the tool definitions
जवाब देखें

जवाब: After the stable instructions and examples

Caching reuses the exact identical prefix.

पाठ: Structuring Prompts for Caching

18. Why validate structured output even when using JSON mode?

  1. Responses can still miss fields, use wrong types or be truncated
  2. JSON mode guarantees business correctness
  3. Validation trains the model
  4. It is only needed for images
जवाब देखें

जवाब: Responses can still miss fields, use wrong types or be truncated

Treat model output like any untrusted input to your code.

पाठ: Structured Output as a Contract

19. Why log the template version with each request?

  1. To hide changes from users
  2. To increase the context window
  3. Because models require it
  4. So results can be reproduced and compared across prompt changes
जवाब देखें

जवाब: So results can be reproduced and compared across prompt changes

Traceability makes prompt changes safe.

पाठ: Versioning Prompts and Context Pipelines

Untrusted Content, Privacy and Grounding

20. Which is the most robust defence against injected instructions in retrieved content?

  1. A longer list of banned phrases
  2. Limiting the model's permissions and requiring confirmation for sensitive actions
  3. Telling the model to be careful once
  4. Using a larger context window
जवाब देखें

जवाब: Limiting the model's permissions and requiring confirmation for sensitive actions

Detection helps, but limiting capability bounds the damage.

पाठ: Prompt Injection Through Context

21. Where should document access rights be enforced?

  1. In an instruction asking the model not to reveal it
  2. At retrieval time, before content reaches the model
  3. After the answer is shown
  4. Nowhere, the model decides
जवाब देखें

जवाब: At retrieval time, before content reaches the model

Context should only contain what the user may see.

पाठ: Data Minimisation in Context

22. Why include unanswerable questions in tests?

  1. Because models cannot cite
  2. To make tests faster
  3. To lower the token count
  4. To check the system admits missing information instead of guessing
जवाब देखें

जवाब: To check the system admits missing information instead of guessing

Honest not-found answers are a quality signal.

पाठ: Grounding, Citations and Saying Not Found

Evaluating and Applying Context Strategies

23. What is an ablation?

  1. Changing or removing one component and measuring the effect
  2. Changing everything at once
  3. Deleting the evaluation set
  4. Retraining the model
जवाब देखें

जवाब: Changing or removing one component and measuring the effect

One variable at a time isolates the cause.

पाठ: Evaluating Context Changes With Ablations

24. In the case study, what fixed citations of outdated policies?

  1. Sending all 3,000 pages
  2. Raising the temperature
  3. Filtering retrieval to current policy versions
  4. Removing the question
जवाब देखें

जवाब: Filtering retrieval to current policy versions

Metadata filtering beats hoping the model notices dates.

पाठ: Case Study: A Support Assistant Over Policy Documents

25. When should you rerun context ablations?

  1. Only when the UI changes
  2. Never, results are permanent
  3. After switching or upgrading the model, and after major prompt changes
  4. Only once a year regardless of changes
जवाब देखें

जवाब: After switching or upgrading the model, and after major prompt changes

Context behaviour depends on the model.

पाठ: A Context Engineering Checklist

← कोर्स पर वापस जाएँ