Coforge & Hexaware: GenAI Interviews

• Fixed-size + overlap (e.g. 512 tokens, 10–15% overlap): simple and a good baseline.
• Recursive character / structure-aware: split on headings → paragraphs → sentences, keeping sections whole. Best for documentation and policies.
• Semantic chunking: split where adjacent-sentence embedding similarity drops. Gives cleaner topics, at extra cost.
• Document-specific: keep tables and code blocks whole; slide-level or page-level for PDFs.
• Parent–child / small-to-big: retrieve small chunks, send the parent section to the LLM.
• Pick by evaluation: measure recall@k and faithfulness for 2–3 strategies on your own Q&A set.

• Hybrid = sparse (BM25 keywords) + dense (embeddings), with results fused using Reciprocal Rank Fusion: score = Σ 1/(k + rank).
• Dense search catches meaning and synonyms. Sparse search catches exact terms: product codes, names, error IDs, acronyms. Embeddings often miss these.
• Add a cross-encoder re-ranker on the fused top 50 for the best precision.
• Available in Azure AI Search, Elasticsearch/OpenSearch, Weaviate, Qdrant and Pinecone (sparse-dense).

1. For each token embedding x, compute Q = xW_Q, K = xW_K, V = xW_V.
2. Scores: QKᵀ, the similarity of each query with every key.
3. Scale by √d_k so the softmax doesn't saturate, then apply a mask (causal or padding).
4. Softmax turns the scores into attention weights that sum to 1.
5. Output = weights × V: each token's new representation mixes information from the relevant tokens. - Cost is O(n²·d) in sequence length, which is why long context is expensive. FlashAttention and sparse or linear attention help.

Answer (practical and specific):

• Engineering productivity: Copilot-style code assistants for tests and boilerplate, always reviewed; LLMs for test-case generation from requirements in QA.
• Frameworks: LangChain/LangGraph for RAG and agents, LlamaIndex for ingestion, RAGAS for evaluation, LangSmith for tracing.
• What I built: "A RAG assistant over 2,000 test-strategy documents. Hybrid search plus a re-ranker; it cut test-planning time by about 30%."
• If challenged: "There are several valid designs. I chose X because of constraints A and B; I'd switch to Y if latency or cost mattered more." Stay calm and evidence-based.

from collections import defaultdict

def group_anagrams(words):
    g = defaultdict(list)
    for w in words:
        g["".join(sorted(w))].append(w)
    return list(g.values())

print(group_anagrams(["eat", "tea", "tan", "ate", "nat", "bat"]))

• O(n · k log k) for k-length words. Using a 26-letter count tuple as the key gives O(n·k).

• Agentic AI: an LLM that plans, calls tools, observes the results and iterates towards a goal, often with memory and several cooperating agents.
• Reliability checklist:
• Few, well-described tools with strict schemas.
• A step and time limit.
• Retries with validation.
• A deterministic workflow where the path is known.
• Human approval for side effects.
• Tracing of every step.
• A trajectory evaluation set.
• Fallback to a human when unsure.