Retrieval and Knowledge Systems

Fixed 512-token chunks, one embedding, top-k, answer: the baseline every later change must beat.

Mine real user questions and label the passages that answer them, instead of generating both.

recall@k, nDCG@10 and MRR, what each one hides, and why answer accuracy will not tell you.

Take one wrong answer and pin it on parsing, chunking, retrieval, ranking, or generation.

Docling and LlamaParse on scanned pages, two-column layouts, and tables that break across pages.

Headings, sections, and parent-child chunks instead of a fixed window that cuts mid-sentence.

Prepend a generated one-line context to each chunk, then measure what it bought and what it cost.

Embed the whole document first and slice after, then compare it against contextual retrieval.

Tenant, source, section path and effective date, designed at ingest rather than bolted on later.

The three content types plain text extraction quietly ruins, and what to store instead.

Qwen3-Embedding, Cohere Embed v4 and Voyage 4 compared on your corpus, not on an MTEB average.

Truncate 3072 down to 512, read the recall you lost, and price the storage you saved.

m, ef_construct and ef: the three numbers trading recall against latency and RAM.

Scalar and binary quantization, and the float32 rescore that often costs more latency than it buys.

Sparse BM25 vectors and dense vectors in one Qdrant query, fused with reciprocal rank fusion.

Payload filters, the filtered HNSW recall cliff, and testing recall under real access rules.

Incremental upserts, tombstones, and re-embedding into a versioned collection before swapping traffic.

Cohere Rerank 4 and Qwen3-Reranker over a hundred candidates, cut to eight, and the latency added.

ColBERT-style multi-vector scoring, sitting between a bi-encoder and a full cross-encoder.

HyDE, multi-query and decomposition, each kept only if it moves recall@20 on your own set.

Order, dedupe and budget the final chunks, because position still decides what the model reads.

Span-level attribution, and the deceptive-grounding case where real evidence names the wrong entity.

Score one fabricated answer with Ragas and DeepEval, then work out why the two disagree.

Run both on the same eval: effective context lands near 50 to 65 percent of the advertised window.

Cost and latency per query for a cached million-token prompt against one retrieval call.

Microsoft GraphRAG, LightRAG and LazyGraphRAG: indexing cost weighed against real multi-hop gain.

Search, verify, refine, stop: capped loops, semantic caching, and a trace for every step.

Evals in CI, drift on the live query log, and a retrieval config you can version and roll back.