A compliance assistant can retrieve the right policy and still give an answer that combines conditions from different clauses. Grounding addresses both sides of this reliability problem: a relevant passage gives the generator evidence to use, and returning that passage with the answer gives a reviewer something to check. A citation is not proof by itself; the cited passage must support the particular claim. section 1.3.5 section 14.6
Keep the path from source to answer inspectable:
- Retrieve relevant evidence before generation. The passage in the prompt can improve accuracy; the passage returned with the answer lets someone verify it. section 1.3.5
- Preserve provenance during extraction. LangExtract attaches character offsets to extracted entities and relations, so a graph fact can be traced back to its source sentence. section 12.1.1
- Check support at the claim level. A retrieved source makes attribution possible, but does not guarantee that every statement in the answer is attributable to it. section 1.3.5
Graph-based RAG makes provenance especially important. A graph assembled from uncited triplets can produce answers that look well connected but cannot be checked at the source. If extraction runs over synthesized wiki pages rather than original documents, offsets must survive both steps: from graph fact to page, then from page to original sentence. Each synthesis step can lose or invent information. This two-step design is a prototype to evaluate on a specific corpus, not an established improvement over chunk-level extraction. section 12.1.1 section 12.6.4
When better retrieval exposes weaker grounding
A compliance assistant initially gave conservative, largely faithful answers while retrieval returned many irrelevant documents. After a reranking fix, context relevance rose from 0.65 to 0.82, but faithfulness fell from 0.91 to 0.84: the generator now made more specific claims, including one that improperly combined policy clauses. The team instructed it to quote exact source language for criteria and thresholds and not combine conditions from different sections; it also lowered temperature from 0.7 to 0.3. Faithfulness recovered to 0.90. Reasoning questions still scored 0.62, so an aggregate score alone would have hidden a remaining failure mode. section 14.6
Table 14.4 Evaluation scores across improvement iterations
|
Iteration
|
Context relevance
|
Faithfulness
|
Answer relevance
|
Primary diagnosis
|
|
Baseline (week 1)
|
0.65
|
0.91
|
0.78
|
FP2: retrieval ranking
|
|
After the reranking fix (week 3)
|
0.82
|
0.84
|
0.81
|
FP3: faithfulness regression
|
|
After the grounding prompt (week 5)
|
0.83
|
0.90
|
0.83
|
Stable: monitor
|
Compare context relevance with faithfulness across the reranking and grounding-prompt changes: better evidence retrieval did not automatically produce better-supported answers. section 14.6












