Workstation Model Services

SignalMind

SI net classification evidence engine. It ranks SI categories for a target net using net name, topology, related nets, current candidate state, exclusions, boosts, and group context.

• Use for: ambiguous net category help inside SI-Extract.
• Output: category rankings, scoring, source agreement, decision flags, retrieval hits.
• Authority: advisory scoring input to the SI-Toolkit pipeline.

ComponentMind

Component evidence engine. It retrieves cited evidence for part numbers, aliases, pin-role evidence, Windchill metadata, and reviewed ComponentLibrary/ComponentMind audit decisions.

• Use for: component lookup, alias repair, cited pin/electrical evidence.
• Output: ranked cited component evidence with allowed-use and exact-match flags.
• Authority: advisory only. It does not replace SignalMind or final SI classification.

SI-Extract LLM

Gemma-4-E4B-it fine-tuned with LoRA SFT on a 24,035-row multi-task corpus generated from the SI-Extract pipeline. Served locally as a Q5_K_M GGUF through llama-cpp-python with an OpenAI-compatible REST shim.

• Use for: advisory verdicts, SignalMind arbitration, ComponentMind grounding, regex-rule proposals, and abstain decisions.
• Output: one JSON object per query, schema gemma4-si-llm-output/v1.
• Authority: advisory only. Smoke-eval JSON parseability was marginal; treat outputs as suggestions, not decisions.

By Service

By Request Type

Top Users

Top Boards

SignalMind Request Fields

• board: board or extraction name.
• net or target_net: target net name.
• related_nets: array of {net, relationship}.
• bus_id, xnet_group, name_family: group context.
• current_category, current_confidence, current_source: current pipeline state.
• observations, pull_info, topology: physical/topology hints.
• physics_eliminated_categories, hard_excluded_categories, soft_excluded_categories, boost_categories, allowed_categories: controls.
• top_k: number of categories/evidence rows to return.

SignalMind Response Fields

• top_category, top_categories, raw_top_categories: ranked SI categories.
• score_details: top scores, margin, dominance, category scores.
• source_agreement: source-level agreement and conflict markers.
• decision: state, confidence, reasons, flags.
• category_evidence: per-category retrieval source support.
• hits: retrieved chunk IDs, categories, source types, and source ranks.
• controls_applied: exclusions, boosts, and allowed-category effects.

ComponentMind Request Fields

• query: direct natural-language query. If present, this is used as-is.
• board, refdes: board context.
• part_number or partNumber: target component part number.
• device, component_class or class: extracted component metadata.
• pin_name or pinName, pin_number or pinNumber: pin evidence lookup.
• net: optional net name tied to the component lookup.
• source_path, observations, context: extra audit context.
• top_k, search_k: returned rows and internal retrieval depth.

ComponentMind Response Fields

• status: ok or no-hit.
• classification_authority: always advisory_only.
• top_results and hits: ranked cited evidence rows.
• exactComponentMatch: true only when the hit matches the requested component key.
• allowedUse: permitted use, such as rag_retrieval, electrical_review, or metadata_only.
• deterministicClassificationAllowed, deterministicPinRoleAllowed: trust flags for downstream policy.
• citation: source path, type, key, and sometimes source hash.
• warnings: guardrail warnings such as no exact component match.

Restart and Logs

Copysystemctl --user status signalmind.service signalmind-tunnel.service
systemctl --user status componentmind.service componentmind-tunnel.service
systemctl --user status siextract.service siextract-tunnel.service

systemctl --user restart signalmind.service signalmind-tunnel.service
systemctl --user restart componentmind.service componentmind-tunnel.service
systemctl --user restart siextract.service siextract-tunnel.service

journalctl --user -u signalmind.service -n 120 --no-pager
journalctl --user -u componentmind.service -n 120 --no-pager
journalctl --user -u siextract.service -n 120 --no-pager

Update Model Assets

Copy# SignalMind runtime
cd /home/elkowalski/regex-augmenter/dist/signalmind_context_engine_runtime
./.venv/bin/python tools/signalmind_runtime.py --json '{"board":"smoke","net":"BMC_I2C_SCL"}'

# ComponentMind runtime
cd /home/elkowalski/componentmind-finetune/dist/componentmind_context_engine_runtime
./.venv/bin/python tools/componentmind_runtime.py --json '{"part_number":"1N4001"}'

# SI-Extract LLM (swap GGUF, then restart)
ls -lh ~/trained-models/model-assets/gguf/siextract-q5km.gguf
systemctl --user restart siextract.service

VPS Proxy Check

Copysudo nginx -t
sudo systemctl reload nginx
ss -ltnp | grep -E ':(8088|8089|8090)\b'

curl -sS http://127.0.0.1:8088/health
curl -sS http://127.0.0.1:8089/health
curl -sS http://127.0.0.1:8090/v1/models

Data Refresh Rules

• Do not hand-edit generated RAG output as a fix. Repair the source data or review-decision generator, then rebuild.
• Keep Windchill metadata as metadata unless reviewed evidence grants stronger authority.
• Preserve citation fields and allowed-use flags in every new evidence row.
• Run stress evals before replacing the deployed model or index.

Near-Term Integration

• Wire the SI-Toolkit ComponentMind client into the component context path with the existing audit-only switch first.
• Use exactComponentMatch as a hard guard before applying component-specific evidence.
• Persist ComponentMind request/response snippets into generated reasoning files for review.
• Compare ComponentMind evidence against existing ComponentKnowledgeLoader behavior.

Dataset and Training

• Expand real-board eval coverage beyond the current stress set.
• Add Windchill refresh jobs that keep metadata separate from reviewed electrical evidence.
• Add negative tests for similar part families, especially logic gates, buffers, connectors, and power parts.
• Re-run the ComponentMind data-readiness audit after each ComponentLibrary or Windchill import.