cai-exos-systems/daveadmin-exos-demo:exosneeds/architecture.md
Current Runtime Pattern
1. The user selects a demo customer and an agent/use case in `demo.exos.bluenotelogic.com`.
2. The UI sends the prompt to the PHP API, which streams Server-Sent Events back to the browser.
3. The backend classifies the prompt into a mode such as quote-to-order, TM Forum mapping,
architecture explanation, or policy/pricing explanation.
4. The tool layer reads BSS state from FOSSBilling: customer, orders, invoices, and support tickets.
5. The catalog layer reads product-offering data from Directus and maps it to TM Forum-style product catalog behavior.
6. The policy layer calls OPA to decide whether the quote/order path is allowed or needs a human gate.
7. The billing/rating layer keeps the FOSSBilling invoice as proof of the price and customer billing state.
8. The standards evidence layer searches TM Forum, EXOS BSS, and telecom reference material.
9. The action layer calls provisioning and entitlement simulators and routes approval/fallout into tickets.
10. The LLM synthesizes the visible trace into a clear answer.
In short:
```text
User prompt
-> intent/mode classifier
-> BSS tools
-> catalog tools
-> OPA policy
-> standards evidence
-> provisioning/ticket tools
-> Qwen synthesis
-> streamed answer + tool trace + audit trail
```
Why The Answer Looks Better Than Plain Chat The answer is grounded by systems and traces: - It can cite the current customer, invoice, ticket, and order state because those came from tools. - It can explain why an action was allowed or blocked because OPA returned a decision. - It can map work to TM Forum APIs because the standards evidence layer and curated TMF map are in the prompt. - It can avoid pretending everything succeeded because tool failures are preserved in the trace. This is the important thesis-to-EXOS bridge: the LLM is not the system of record and does not own the action. It reads a governed trace and explains the outcome.
EXOS Baseline For The Next Build Start with the simplest credible EXOS baseline: - Qwen 7B only - no tuned LLM - no RAG in the first pass - tool traces still visible - OPA decisions still visible - BSS and TM Forum API mappings still structured Then add the thesis evidence pattern: - index TM Forum/ODA/Open API docs, EXOS architecture docs, runbooks, product catalog notes, and demo scenarios - retrieve the top evidence chunks per question - pin the evidence packet for the LLM call - score whether the answer is supported by the retrieved evidence - store every run as an audit artifact EXOS Naming Use EXOS language: - EXOS standards evidence layer - EXOS controlled evidence layer - EXOS validation harness - EXOS agent audit trail Avoid calling the EXOS version CaveauAI in client-facing materials. ```
Why The Answer Looks Better Than Plain Chat The answer is grounded by systems and traces: - It can cite the current customer, invoice, ticket, and order state because those came from tools. - It can explain why an action was allowed or blocked because OPA returned a decision. - It can map work to TM Forum APIs because the standards evidence layer and curated TMF map are in the prompt. - It can avoid pretending everything succeeded because tool failures are preserved in the trace. This is the important thesis-to-EXOS bridge: the LLM is not the system of record and does not own the action. It reads a governed trace and explains the outcome.
EXOS Baseline For The Next Build Start with the simplest credible EXOS baseline: - Qwen 7B only - no tuned LLM - no RAG in the first pass - tool traces still visible - OPA decisions still visible - BSS and TM Forum API mappings still structured Then add the thesis evidence pattern: - index TM Forum/ODA/Open API docs, EXOS architecture docs, runbooks, product catalog notes, and demo scenarios - retrieve the top evidence chunks per question - pin the evidence packet for the LLM call - score whether the answer is supported by the retrieved evidence - store every run as an audit artifact EXOS Naming Use EXOS language: - EXOS standards evidence layer - EXOS controlled evidence layer - EXOS validation harness - EXOS agent audit trail Avoid calling the EXOS version CaveauAI in client-facing materials. ```