Original OrbTrail analysis expanded with complementary research, practical context and verified references.
Salesforce published a set of tests for improving RAG answers in Data 360 on September 2, 2026. On the WixQA benchmark, a configuration package raised accuracy from 62.5% to 92.5% without custom code or model training. On complex enterprise documents from OHRBench, the pipeline reached 86.8% end-to-end accuracy. These are substantial, reproducible results on public datasets, but they are not a promise that every help center, technical manual or internal policy repository will gain the same thirty percentage points.
The central question is therefore more precise than 'which setting improves RAG?': how can a team tell whether an error starts in file processing, content splitting, retrieval or generation—and prove that a fix transfers to its own corpus before expanding context? Salesforce's measurements help separate the layers. After full optimization, 40% of remaining errors came from generation, 26% from content processing, 18% from search and 16% from splitting. That distribution belongs to the reported tests, and Salesforce explicitly says it will vary by document type.
OrbTrail's analysis is that the most important development is not one accuracy figure but the opportunity to make RAG diagnosable. Replacing the model because an answer is wrong wastes effort when a table was destroyed at ingestion. Increasing top-k can add noise when the correct passage was already near the top. A stricter prompt cannot recover a diagram that never entered the index. A defensible rollout begins with answerable questions and changes one layer at a time while holding the model, corpus and scoring method steady.
The page must survive ingestion
On OHRBench as reported by Salesforce, intelligent parsing moved end-to-end results from 65% to 84.4%; the feature is expected in November 2026.
Keep the procedure with its warning
Chunk size should preserve the unit needed to answer, rather than satisfy a fixed token count in isolation.
Relevance beats volume
Enriched indexing added 2.4 points in the cited test; higher top-k helps only when it contributes evidence instead of distractors.
Correct context can still yield a wrong answer
Salesforce attributed 40% of residual errors to generation in its benchmarks; faithfulness and completeness need their own measures.
The local corpus makes the decision
Golden questions, expected answers, supporting passages and layer-level scores form the gate before promotion.
Intelligent parsing repairs the page before search interprets it
The first complementary question is where information can disappear before retrieval even begins. OHRBench contains 8,561 document images from seven domains and 8,498 question-answer pairs derived from multimodal document elements. Its authors show that semantic and formatting noise introduced by OCR cascade through the rest of a RAG pipeline. A shifted table cell, incorrect reading order or chart reduced to incomplete text is not merely a display defect: it changes the evidence that search can retrieve from the evidence a person sees on the published page.
In Salesforce's tests, intelligent parsing evaluates each page and routes predominantly textual or visual material to different extraction methods. On the reported OHRBench slice, accuracy for charts and diagrams rose from 57.9% to 76.3%; reading order across multicolumn or multipage layouts rose from 11.9% to 67.1%; and end-to-end accuracy moved from 65% to 84.4%. Salesforce says the feature is expected in November 2026. It therefore belongs in a roadmap as stated future availability, not as a capability already guaranteed to every organization or region.
The practical decision is to classify the corpus before selecting a remedy. Separate simple HTML articles, text-first PDFs, documents dominated by tables, slide decks, diagram-heavy manuals and scanned files. For each class, choose questions whose answers depend on that exact structure: a table row and column, the complete order of a procedure, or a condition in a side note. Compare extracted content with the source page before scoring an agent answer. If evidence is already corrupted at this stage, changing embeddings, prompts or models only conceals the defect.
A benchmark measures the corpus it received—not yours
The second question is what 92.5% actually proves. WixQA was built around a snapshot of Wix's public help knowledge base and contains three datasets: 200 real user queries with expert-written multistep answers, 200 dialogue-derived pairs validated by experts, and 6,222 synthetic pairs, one systematically generated from each knowledge article. This is much closer to enterprise support than generic encyclopedia trivia. It still reflects the vocabulary, editorial structure, language and policies of one product, rather than automatically measuring legal contracts, field manuals or internal knowledge written in Portuguese and Spanish.
Salesforce reports that, on the same benchmark, moving from 512- to 2,048-token chunks, increasing retrieval depth from 10 to 30 passages and applying a domain-tuned generation prompt raised accuracy from 62.5% to 92.5%. Those changes were evaluated as the stated configuration. Teams should not turn 2,048 and 30 into universal defaults: a long procedure can benefit from continuity, while a repository of short, repetitive articles may produce large chunks full of competing subjects. The useful number is the improvement observed under controlled conditions on the local corpus.
An independent warning comes from Lost in the Middle. The study found that models could use evidence less effectively when it appeared in the middle of long contexts and, in an open-domain QA experiment, performance saturated before retriever recall did. Its models and dataset date from 2023, so the result does not quantify today's Data 360 models. It supports a still-useful architectural precaution: more retrieved material creates opportunity and distraction at the same time. Test top-k as a curve—perhaps 5, 10, 20 and 30—and record recall, faithfulness, latency and consumption instead of approving the largest value by default.
A larger top-k enters production only with layer-level evidence
The third question is how to convert the demonstration into an operating gate. Start with 20 to 50 questions paired with correct answers, the range Salesforce recommends for local evaluation. The sample must represent actual work rather than easy wins: include single-page answers, multistep procedures, tables, questions that have no answer in the corpus, conflicting versions and customer vocabulary that does not literally appear in the source. Store the authoritative page and passage for each question. Without that evidence, a final score cannot reveal whether the system arrived at a correct answer for the right reason.
Then score the layers separately. The Ragas research framework distinguishes whether retrieval finds relevant, focused passages, whether the model uses those passages faithfully and whether the generated answer addresses the question. An organization does not have to adopt a specific library to use that decomposition. It can measure hit or recall for the expected passage, precision of retrieved context, claim-level faithfulness, answer completeness and correct refusal when the repository lacks support. Human review remains necessary for high-impact cases and to calibrate any automated evaluator.
A sequential experiment prevents false conclusions. First freeze the model and prompt while comparing parsing; then freeze the best extraction and vary chunking and indexing; only afterward compare top-k and generation prompts. Promote a configuration when it improves the complete set without an unacceptable regression in the most critical class, keeps unsupported answers below the defined threshold and fits latency and consumption budgets. Rerun the suite after material corpus changes. OrbTrail's conclusion is straightforward: Salesforce shows that configuration can recover a great deal of performance, but the durable asset is not the winning configuration—it is the method that explains why it won and detects when it stops winning.
