AI Engineer (LLM & Generative AI)
RAG and vector search questions
181 questions on RAG and vector search for ai engineer (llm & generative ai)s, easiest first.
Easy 52
- What text embeddings areeasy
- What RAG doeseasy
- Comparing two embeddingseasy
- Dot product of two vectorseasy
- What a vector database doeseasy
- Which words BM25 rewardseasy
- What semantic search adds over keywordseasy
- What an embedding dimension iseasy
- What top-k means in retrievaleasy
- Why vector search is approximateeasy
- When a flat index is fineeasy
- Why store metadata with each chunkeasy
- Choosing pgvectoreasy
- Long chunks and embedding limitseasy
- Arabic questions, English documentseasy
- Why bi-encoders scale to big corporaeasy
- Handling an edited source documenteasy
- Sparse versus dense vectorseasy
- Bot quotes a withdrawn documenteasy
- Reading distance scoreseasy
- When document and query vectors are madeeasy
- Why documents are split into chunkseasy
- Why chunks overlapeasy
- How a recursive text splitter workseasy
- Chunks that lost their headingeasy
- Tables in PDFs come out garbledeasy
- Scanned PDFs return no texteasy
- What hybrid search meanseasy
- Rewriting the user queryeasy
- Restricting search to the right yeareasy
- Matching "refunds" and "refunded"easy
- Order of a RAG ingestion pipelineeasy
- Cleaning scraped web pageseasy
- Searching with several phrasingseasy
- One chunk is not enougheasy
- What semantic chunking doeseasy
- Why RAG answers show their sourceseasy
- Computing precision at 5easy
- What retrieval hit rate measureseasy
- What a golden eval set containseasy
- What faithfulness measureseasy
- What answer relevance measureseasy
- Keeping salary documents privateeasy
- Personal data in the indexeasy
- The slowest step in a RAG requesteasy
- What to log for each RAG requesteasy
- Using thumbs-up and thumbs-downeasy
- A document that gives orderseasy
- When you may not need retrievaleasy
- What LLM-as-a-judge meanseasy
- What the embedding layer doeseasy
- What CLIP learnseasy
Medium 76
- Choosing a chunking strategymedium
- Knowledge that changes every daymedium
- Embeddings miss exact product codesmedium
- Adding a rerankermedium
- Cosine similarity by handmedium
- Computing recall at 5medium
- Computing MRRmedium
- Counting chunks with overlapmedium
- Keeping tenants apart in RAGmedium
- Tuning HNSW search breadthmedium
- Reciprocal rank fusion scoresmedium
- Dot product on unit vectorsmedium
- Cosine of a scaled vectormedium
- Semantic cache gives wrong answersmedium
- Memory for one million float vectorsmedium
- Memory after int8 quantisationmedium
- Binary quantisation with rescoringmedium
- Truncating Matryoshka embeddingsmedium
- L2 versus cosine on unit vectorsmedium
- Squared distance from a cosine scoremedium
- Raising nprobe in an IVF indexmedium
- Raising M when building HNSWmedium
- Forgetting the query prefixmedium
- Short questions against long passagesmedium
- Insert fails with a dimension errormedium
- How product quantisation saves memorymedium
- Fine-tuning embeddings for your domainmedium
- Batch size in contrastive trainingmedium
- How ColBERT scores documentsmedium
- What SPLADE adds to keyword searchmedium
- Cost to embed a document corpusmedium
- Shards versus replicas in a vector DBmedium
- Index slows after many deletesmedium
- Picking an embedding modelmedium
- What k1 controls in BM25medium
- Setting BM25 b to zeromedium
- Working out an IDF weightmedium
- How HyDE searchesmedium
- Splitting a compound questionmedium
- Retrieve small, return the parentmedium
- Adding document context to each chunkmedium
- What maximal marginal relevance doesmedium
- Turning words into metadata filtersmedium
- Routing questions to different sourcesmedium
- Chunking well-structured documentationmedium
- Splitting a long tablemedium
- Chunking a code repositorymedium
- Favouring newer documentsmedium
- Reranking time for 100 candidatesmedium
- Using an LLM as a rerankermedium
- What late chunking doesmedium
- Making nightly re-indexing cheapermedium
- Queries full of local acronymsmedium
- Letting the model search againmedium
- What context recall measuresmedium
- What context precision rewardsmedium
- Eval questions generated from chunksmedium
- Judge prefers whichever answer is firstmedium
- Can you trust the LLM judge?medium
- Offline versus online RAG evaluationmedium
- Small change, broken answersmedium
- Adding up time to first tokenmedium
- Trimming retrieved chunksmedium
- Two sources that disagreemedium
- Cost of one RAG answermedium
- A deletion request for indexed datamedium
- Spotting gaps in the knowledge basemedium
- The bot says "I don't know" too oftenmedium
- Citations users can actually checkmedium
- Arabic question, English sourcesmedium
- Measuring a support bot in productionmedium
- Building an eval set from logsmedium
- Parameters in the embedding matrixmedium
- Tying input and output embeddingsmedium
- Positives and negatives in CLIP trainingmedium
- Pooling a decoder model into one vectormedium
Hard 53
- Stuffing 100 chunks into the prompthard
- Upgrading the embedding modelhard
- Computing nDCG at 3hard
- Budgeting the context windowhard
- How many results to rerankhard
- Adding BM25 and cosine scoreshard
- IVF index after a new languagehard
- A filter that matches 0.1% of datahard
- All similarity scores look highhard
- One similarity cut-off for all querieshard
- One chunk in every result listhard
- How many candidates to rescorehard
- Embeddings and negated statementshard
- Duplicates during contrastive traininghard
- Hinglish queries in Latin scripthard
- Inner product with unnormalised vectorshard
- Storage cost of token-level vectorshard
- Choosing a vector index configurationhard
- One vector for a 40-page PDFhard
- A billion vectors on a budgethard
- Top results are copies of one pagehard
- Answer split across two chunkshard
- When HyDE makes retrieval worsehard
- Questions about the whole corpushard
- Garbled chunks from two-column PDFshard
- Cost of context notes per chunkhard
- Reranker misses facts at the endhard
- MMR set too far towards diversityhard
- Rewritten queries change the meaninghard
- Setting the balance in hybrid searchhard
- A question that needs two hopshard
- Relative dates break extracted filtershard
- Re-runs create duplicate chunkshard
- Parent retrieval floods the prompthard
- RAG gets totals wronghard
- Generating questions for each chunkhard
- Moved teams, still sees old documentshard
- Leaking data through an image linkhard
- Cached answers cross a permission linehard
- Perfect faithfulness, unhappy usershard
- Suspiciously high eval scoreshard
- Whole corpus in context versus RAGhard
- A 4-point gain on 50 questionshard
- Cache serves an outdated policyhard
- False facts planted in a wikihard
- Citations to chunks that do not existhard
- Better offline, flat onlinehard
- Data that must stay in the countryhard
- List questions return partial listshard
- Faithful answer, wrong documenthard
- Tokens that make the model behave oddlyhard
- Teaching new facts by fine-tuninghard
- Adding domain tokens to a tokenizerhard