Ask the corpus
The classifier routes each question. Document searches use only the selected manual, with citations to the source PDF.
Classifier routed response
The classifier chooses a predefined reply or routes the question to document retrieval.
Ask a question to see the routed response here.
Direct RAG response
Forces document retrieval on the selected manual, bypassing classifier routing.
Use “Compare classifier vs RAG” to run the comparison.
Documents
Indexed versions available to retrieval.
| Title | Pages | Extraction | Exclusions | Version |
|---|
Artifacts
Parsed Markdown and JSON written after ingest.
Select extracted.md or extracted.json
Upload a document and chunk map
Upload one PDF with its Excel chunk map. The map controls the page ranges and boundaries used to create searchable chunks.
Reference files: UG manual PDF · UG chunk map Excel
Ingestion queue
| State | File | Job |
|---|
Create a chat widget
Generate an embed for one website domain or IP address. Requests are restricted to the value you enter.
Configured widgets
Disable a widget to immediately stop requests using its key.
RAG evaluation
Create a grounded question set, run it through retrieval, and inspect judge scores with the full retrieval trace. Dataset generation sends the extracted document text to the selected provider.
1 · Build a question dataset
The example fills in the selected document IDs. Replace its question and answer with facts supported by that document before importing; claims, quotes, pages, and tags are optional.
2 · Choose the judge
Evaluation reports section hit rate at Top-1, Top-2, and Top-3 when expected_sections are supplied, plus judge-estimated evidence relevance and answer coverage.
Evaluation progress
Saved runs
Dataset questions
How to read these scores
Scores are judge estimates, not exact ground-truth labels. Treat them as signals to inspect. Below 0.60 is flagged for improvement; 0.60–0.79 is a review range; 0.80+ is generally strong. These are practical UI thresholds, not universal benchmarks.
| Measure | Plain-language meaning | Technical meaning | Stages to inspect if low |
|---|
Run report
Select a saved run to view aggregate scores and per-question trace.
Question trace
Click a question in the report to inspect stage results and retrieved passages.
Releases
Activate a built release to serve it for answers and retrieval.
| Release | State | Docs |
|---|
Recent answer requests
| Time | Status | Question | Request |
|---|
Session
Admin APIs use your secure, HttpOnly sign-in session.
Last API payload
Inspector for the most recent JSON response.
{}