What are you seeing?

First check

Look at the parser output for that table before blaming retrieval.

Read

• RAG-08Parsing, layout and OCR
• RAG-09Tables, images and mixed content
• RAG-14Structure aware chunking and tables

First check

Find where the answer passage was split across chunks.

Read

• RAG-14Structure aware chunking and tables
• RAG-15Chunk size, overlap and token based sizing
• RAG-17Parent child and multi representation indexing
• RAG-16Contextual chunking and late chunking

First check

Run a keyword (BM25) search for the code on its own.

Read

• RAG-27Dense retrieval and keyword search with BM25
• RAG-28Hybrid retrieval and rank fusion

First check

Compare the rank before and after reranking.

Read

• RAG-31Reranking: two stage, cross encoder and LLM rerankers

First check

Log the rewritten query the system actually searched with.

Read

• RAG-26Query understanding, rewriting and expansion
• RAG-38Conversational RAG, answer UX and RAG APIs

First check

Check that the right chunk was really in the prompt.

Read

• RAG-34Grounding prompts and abstention
• RAG-35Lost in the middle and context management

First check

Check the document version and the indexed date of the chunk used.

Read

• RAG-12Updates, deletes, reindexing and ingestion pipelines
• RAG-36Conflicts, staleness and freshness rules
• RAG-47Index lifecycle operations

First check

Check that the access filter ran on every query, not only some.

Read

• RAG-29topK, no good match, and metadata filtering
• RAG-46Access control, privacy and governance

First check

Time each stage separately: retrieval, reranking, generation.

Read

• RAG-49Latency, cost, caching and model routing
• RAG-30Routing, diversity, deduplication and caching

First check

Count tokens per query and see where they go.

Read

• RAG-49Latency, cost, caching and model routing
• RAG-33Context selection, packing and citations

First check

Build a small golden set before changing anything else.

Read

• RAG-39Why measure, and what to measure
• RAG-40Building a golden dataset
• RAG-41Retrieval metrics: precision, recall, MRR, nDCG
• RAG-44Offline and online evals, regression tests and slices