Information Extraction: AI-103 Domain 3 (10-15%)

Nothing is retrievable until it is ingested

Before anything can be grounded, content has to be extracted, chunked, embedded and indexed. That pipeline is the prerequisite, which is why the answer to "we have ten thousand internal PDFs" is an ingestion and indexing pipeline rather than a larger context window or a fine-tuned model. OCR is the first step for anything scanned: without it a PDF is an image and nothing downstream has text to work with.

Layout is meaning

A plain left-to-right sweep across a two-column report reads across both columns, interleaving unrelated sentences into text that is syntactically fine and semantically nonsense. Layout analysis recovers the intended reading order before anything is chunked. The same principle extends to producing clean markdown or structured output before indexing: headings, tables and lists carry relationships, and a table flattened into a run-on line has lost the relationships between its cells.

Chunking is a precision trade-off

Chunks that are too large span several topics, so they match many queries weakly and none strongly while spending context on material the query did not need. Chunks that are too small lose the context that made them interpretable. The durable practice is to split on semantic boundaries and carry heading context into each chunk, so a retrieved passage arrives knowing what section it came from.

Semantic, keyword and hybrid

Vector search captures meaning and handles paraphrase, which is what makes conceptual questions work. It is weakest on exact identifiers — product codes, error numbers, names — because a code carries little semantic content and its nearest neighbours in vector space are other codes. Keyword matching finds those exactly. Hybrid search runs both and combines them, which is why it is the usual default. Semantic ranking then reorders candidates by how well they answer the query's intent rather than by term overlap.

Field extraction beats pattern matching

For invoices, forms and similar documents, a model-based extraction analyzer that returns a defined schema generalises across layouts in a way that a regular expression per field does not — the first supplier who moves the total breaks the regex. Enrichment during ingestion adds derived structure such as entities, captions and detected language as filterable fields, so that analysis happens once rather than on every query.

Security and provenance are ingestion decisions

Access control belongs at retrieval: filter on a trusted security or department field so restricted content never enters the model's context, because anything the model sees it can leak. A prompt instruction not to reveal other departments' documents is not an access control. Citation is the same kind of decision — if each chunk carries its document identifier and location, the answer can point back to it; retrofitting provenance after generation is guesswork.

Exam tip

When retrieval is returning the wrong thing, work backwards through the pipeline rather than reaching for a bigger model. Most retrieval failures on this exam are chunking, layout or indexer-freshness problems wearing a generation costume.