AI-103 information extraction solutions explained
Implement information extraction solutions is worth 10–15% of AI-103, and despite the modest weight it is the domain that quietly decides whether the rest of your solution works. It covers two things: building the retrieval and grounding pipeline that feeds a RAG application, and pulling structured content out of documents. Every agent that answers from your own data depends on both.
Retrieval and grounding pipelines
The objectives walk through a pipeline end to end.
Ingest and index content — documents, images, audio and video. Content is not limited to text files, and a scenario mentioning scanned PDFs, recordings or screenshots is still an ingestion question.
Configure the search strategy. Three options appear, and you need to know what each is good at:
| Strategy | Matches on | Strong when |
|---|---|---|
| Keyword | Exact terms | Product codes, names, identifiers |
| Vector | Meaning | The user’s words differ from the document’s |
| Hybrid | Both, combined | Real-world mixed queries — usually the safe answer |
| Semantic ranking | Re-orders results by meaning | Recall is fine but the best result is not at the top |
Hybrid search is the pragmatic default in most scenarios, and semantic ranking is the answer when the correct document is being retrieved but is ranked too low.
Enrichment with custom or built-in skills for text, images and layout, so that the indexed representation is richer than the raw file.
RAG ingestion flow, including OCR for documents that are images of text rather than text.
Connecting retrieval directly to workflows and agent tools, so the index is something the agent queries, not a separate application feature.
Extracting content from documents
The second group is about turning documents into something a model can use:
- Multimodal pipelines that combine OCR, layout analysis and field extraction. OCR alone gives you characters. Layout analysis tells you what is a table, a heading or a key-value pair. Field extraction returns named fields. A question that asks for an invoice total as a value, not as text somewhere on a page, is asking for all three.
- Clean, grounded representations for agents and RAG using Content Understanding.
- Analyzers that emit structured or markdown output for downstream reasoning.
Markdown output deserves a mention. Handing a model a markdown version of a document preserves headings and tables, which improves both chunking and the model’s ability to reason about structure — a plain text dump loses exactly the signal the model needs.
Where these questions usually go wrong
Two traps recur. The first is choosing a strategy that cannot possibly match the query: vector search for an exact part number, or keyword search for a paraphrased question. The second is answering a structure problem with a text tool — OCR where field extraction is required, or raw text chunks where markdown would have preserved the table.
Sample questions
Question 1. Engineers search an internal knowledge base using both exact part numbers such as 'FLT-2291' and natural questions such as 'why does the filter clog in cold weather'. Which retrieval configuration serves both?
- A. Keyword search only
- B. Vector search only
- C. Hybrid search combining keyword and vector retrieval
- D. Keyword search with the number of returned results increased to fifty
Show answer
Answer: C
Hybrid search combines keyword matching, which finds exact identifiers, with vector matching, which handles paraphrased questions. Keyword alone fails the natural language query, vector alone is unreliable for exact codes, and increasing the result count does not fix a strategy that cannot match the term at all.
Question 2. Ten years of scanned purchase orders must become records containing supplier, order number, date and line items, ready for an agent to query. What pipeline do you build?
- A. A pipeline combining OCR, layout analysis and field extraction
- B. OCR only, then full-text keyword indexing
- C. Image captioning on each scanned page
- D. Manual data entry into a spreadsheet
Show answer
Answer: A
Scanned pages need OCR to produce text, layout analysis to understand tables and key-value regions, and field extraction to return named fields. OCR on its own returns unstructured characters, keyword indexing of raw text does not produce fields, and image captioning describes the page rather than extracting data.
Question 3. Your index reliably contains the right document, but users say the best answer often appears fourth or fifth in the results. Recall is good; ordering is not. What is the most targeted fix?
- A. Re-index the entire corpus with a smaller chunk size
- B. Enable semantic ranking over the existing results
- C. Return twenty results instead of five
- D. Switch the index to keyword-only search
Show answer
Answer: B
Semantic ranking re-orders an existing result set by meaning, which is exactly the stated problem. Re-indexing everything with smaller chunks is a larger change aimed at recall, adding more results makes the ordering problem worse, and switching to keyword-only search would reduce recall.
What to practise
Index a folder of your own documents, including at least one scanned PDF, and query it three ways: keyword, vector and hybrid. Use a query with an exact code and a query phrased as a question, and note where each strategy fails. Then extract fields from the scanned PDF and compare that output with what plain OCR gives you. The difference between those two outputs is the whole domain in one screen.