Evidence Explorer: information flow
How information moves from partner reports to the Evidence Explorer app, and where AI models are involved. Numbers on the diagram match the table below.
Information flows
| Flow | What moves | AI model | |
|---|---|---|---|
| 1 | Partner organisations → Salesforce | Project reports: progress, final and evaluation reports. | No |
| 2 | Salesforce and VicHealth teams → Landing folder | Report files and Impact Logs (PDF, Word, Excel, PowerPoint), uploaded as files. There is no automated Salesforce feed into the pipeline. | No |
| 3 | Landing folder → Bronze (Ingest) | Each new file, stored unchanged along with its file details. Duplicate files are skipped. | No |
| 4 | Bronze → Silver (Extract text) | Plain text extracted from each file and split into sections. | No |
| 5 | Silver → AI model → Gold (Extract evidence) | Report text is sent to Claude Sonnet 4.6. Structured evidence items come back and are saved to Gold as “pending review”. | Yes |
| 6 | Gold ↔ AI models (Match shifts) | Evidence descriptions are sent to the embedding model and to Claude, to group related evidence into shifts. The results are saved to Gold. | Yes |
| 7 | Gold ↔ Evidence Explorer | The app reads evidence and shifts. It writes back review decisions, outcome assignments, an audit log and chat history. | No |
| 8 | Bronze → Evidence Explorer | The original file, when a user opens the source document behind an evidence item. | No |
| 9 | Evidence Explorer ↔ AI models (Chat) | The user’s question and relevant approved evidence are sent to the models. The answer is shown to the user and saved in chat history. | Yes |
| 10 | VicHealth staff ↔ Evidence Explorer | Staff sign in with Entra ID to browse evidence, review it and use the chat. | No |
Where processing happens One item to confirm
Everything runs in the Australia East workspace except the AI models, which Databricks hosts. Of the two models, one is processed within Australia and one can be processed outside it.
| Component | Used in | Where it is processed | Basis |
|---|---|---|---|
| Tables, pipeline job and app | All flows | Australia East | The workspace region, confirmed by VicHealth during discovery. |
| Claude Sonnet 4.6 (language model) | Flows 5, 6 and 9 | Within Australia | Microsoft lists it as available to Australia East workspaces without cross-geography routing. |
| GTE Large (embedding model) | Flows 6 and 9 | Can be outside Australia | Microsoft lists it as needing cross-geography routing in Australia East. |
- What the embedding model receives. Evidence descriptions, which are the short AI-written summaries of each evidence item (flow 6), and the questions users type into the chat (flow 9). It does not receive the original files or the full report text.
- Processed, not stored. Microsoft states that content is not stored outside the workspace’s geography, whatever the cross-geography setting.
- Geography, not region. For Claude Sonnet 4.6 the commitment is processing within the Australia geography, which is broader than the Australia East region.
- To confirm. The workspace’s cross-geography setting has not been checked directly. The Databricks administrator should confirm it.
- Option to keep all AI processing in Australia. Microsoft lists another embedding model, Qwen3 Embedding 0.6B, as available to Australia East workspaces without cross-geography routing. Switching would mean re-creating the embeddings for existing evidence and shifts, and re-tuning the shift matching.
References: model availability by region and Databricks Designated Services.
How the AI models are used
The models do not read the tables and are not a validation step. The pipeline job and the app each send text to a model and receive a response, one request at a time. Results are saved only to the Gold tables.
No AI is involved in moving files into Bronze or extracting text into Silver.
How validation works
People validate the evidence, not the model. Every evidence item arrives in Gold as “pending review”, and reviewers approve or reject it in Evidence Explorer.
The chat and the outcome roll-ups only use approved evidence.