1
Query preparation
The browser builds a compact request with normalized text, intent hints, language context, and optional ML signals.
Purpose
Conversation UI, evidence-first backend
The interface is chat-like, but the runtime is not a generic chatbot. It is a staged technical workflow designed to keep answers inspectable, bounded by evidence, and stable enough for engineering and operational use.
Primary mode
Dataset-first retrieval
Language
Prompt in, answer out
On failure
Local fallback, never silent
Stage 1
Intent & Routing
The browser validates scope, normalizes the query, detects language, extracts terms, and may attach frontend ML signals. This keeps the backend payload small and targeted.
- English and Turkish are first-class inputs.
- Frontend ML assists ranking when available, but the system works without it.
- Unsupported prompts are rejected before retrieval.
Stage 2
Execution Flow
2
Dataset retrieval
The backend searches versioned local datasets first. Scoring favors the user's language for relevance.
3
Context enrichment
If the query is weakly matched or asks for broader context, NQ adds Wikipedia and curated industrial RSS.
- Wikipedia fills neutral background gaps.
- RSS is limited to approved industrial, manufacturing, and IIoT sources.
4
Evidence packet
The strongest dataset, Wikipedia, and RSS excerpts are compacted into a single reference packet. The answer layer only sees that packet.
5
Synthesis
The LLM synthesizes the answer from the evidence packet. Helper models pre-analyze source relevance. If the remote path is unavailable, NQ returns a local evidence summary.
6
Inspectable delivery
The response arrives with runtime metadata: evidence, tools used, model status, and rate limits — all inspectable.
Stage 3
Evidence Sources
Local datasets
Versioned JSON datasets for industrial automation, protocols, process control, SQL, and ML fundamentals.
Wikipedia
Neutral background context when local datasets mention a concept but don't fully unpack it.
Approved RSS feeds
Restricted to operational sources: Automation World, ISA Interchange, MachineMetrics, Robohub, and select manufacturing blogs.
Fallback summaries
If the remote LLM is unavailable, the system formats a local response from retrieved evidence instead of failing silently.
RSS discovery hubs are not used directly. Only concrete feed endpoints enter the runtime list.
Stage 4
Runtime Visibility
- The thread exposes which tools were involved: datasets, Wikipedia, RSS, and the answer layer.
- Model status is inspectable, including verification outcome and fallback use.
- Evidence cards stay tied to the answer.
- Rate-limit info is visible when quotas become relevant.
Stage 5
Safeguards
- Requests are validated and rate-limited before retrieval begins.
- The answer layer stays inside provided references.
- Remote failure triggers a local fallback instead of a dead end.
- Runtime status is subordinate but available — failures are diagnosable without cluttering the thread.
A conversational workspace on the surface. A constrained, inspectable retrieval pipeline underneath.