10. GraphRAG: the PDFs join the graph

Section 1

Page 34

Next section

Section 2

What vector search is good at, and what it is not

Next section

Section 3

Getting documents in

Next section

Section 4

Building the lexical graph

Next section

Section 5

Retrieval, in six steps

Next section

Section 6

Two graphs, on purpose

Next section

Section 7

Extraction that knows what to look for

Hands on

Make the credit policy queryable

40 min

Chunk the Meridian Credit Policy PDF, extract concepts, build a small lexical graph in Fuseki, embed the chunks into a local vector index, then run a hybrid retrieval that answers the arrears question with a page citation and links the answer to the ontology class.

make lab-up            # once, starts Postgres + Fuseki + Ontop
make lab lab10-graphrag

Full instructions, starter files and the solution are in labs/modules/lab10-graphrag/README.md. The lab checks its own work, so you will know if it passed.

Check yourself

Quiz

7 questions, pass mark 70%. Answers are explained either way, so a wrong guess is not wasted.

1 Vector-only RAG answers "what does the credit policy say about arrears?" well but fails on "which of Daniel's documents support the income figure the assessment used?". Why?

The second question is longer, so the embedding is less precise The second question is about a chain of relationships, and an embedding encodes what a passage is about, not how it connects to other passages The payslip was never indexed The second question needs a larger k value

2 Which two things does the two-tier split between the lexical graph and the curated semantic layer graph give you? (pick two)

Select all that apply.

Low-confidence extraction candidates are kept for review and quality analysis rather than discarded The curated graph answers business questions using only validated, ontology-aligned facts The lexical graph can be queried with SQL while the curated graph uses SPARQL Extraction confidence scores are stored on the curated facts so callers can weigh them

3 Why must the model that embeds an incoming query be the same model and version used to build the index?

Different models produce different dimension counts, so the search throws an error Each model defines its own coordinate space, so cross-model cosine similarity still returns numbers but they are meaningless and no error is raised The index stores the model name and refuses queries from other models Query embedding is cached per model, so a mismatch is only a performance problem

4 The completion handler publishes to an SNS topic that four SQS queues subscribe to, instead of invoking the four services directly. Which is not a reason given for that?

The subscribers scale independently, so slow embedding work does not hold up fast index writes A failure in one subscriber does not affect the other three SQS retention lets you re-drive one subscriber after fixing a bug, without re-running the PDF parse SNS guarantees the four subscribers process the document in a fixed order

5 Which edge in the lexical graph is what makes a GraphRAG answer able to cite "page 34"?

The concept-to-concept relationship edge The concept-to-chunk edge, since each retrieved concept knows which chunk mentioned it and each chunk carries its page number The vector similarity score stored on the concept node The chunk-to-document edge alone

6 An LLM extracts Daniel Okafor's annual income as 54,000 at confidence 0.91 and his employer as a payroll provider name at 0.62. Meridian's shape requires 0.75 minimum. What happens?

Both are promoted to the curated graph, with the confidence scores attached The income is promoted; the employer name stays in the lexical graph where an exploratory query and a human reviewer can find it Both are rejected, because one failing constraint fails the whole extraction The employer name is deleted, since unvalidated data is a compliance risk

7 You need to pull loan IDs and write-off dates out of 40,000 loan agreements. Which technique does the guidance point you at first?

An LLM via Bedrock, since it understands context best spaCy with rules and patterns, because these are well-defined entities with regular surface forms and it is fast, cheap and deterministic A custom fine-tuned model trained on the agreements Amazon Textract, because every document needs OCR