LLM And Artifacts

This page explains when AutoBiology uses an LLM, what remains deterministic, and how to use the repository artifacts published on GitHub.

Check Whether LLM Is Enabled

Run:

autob config show

LLM inference can run only when all of these fields are set:

• apiKey: shown as redacted instead of unset
• baseUrl: not unset
• model: not unset

After a pipeline run, check:

• run-meta.json: config.llmModel is the configured model name, or not-configured when no LLM client was created.
• 06-clarifications.json: contains an LLM 辅助层未启用 clarification when the LLM layer was unavailable or failed.
• 04-requirements.json: LLM-added requirements use inferenceRule: "LLM-Candidate" and status: "candidate".

Copyable checks:

node -e "const fs=require('fs'); const m=JSON.parse(fs.readFileSync('out/run-meta.json','utf8')); console.log(m.config.llmModel)"
rg -n '"LLM-Candidate"' out/04-requirements.json
rg -n 'LLM 辅助层未启用' out/06-clarifications.json

Default Behavior

By default, AutoBiology does not call an LLM unless the JSON config resolves to a complete OpenAI-compatible client. The deterministic pipeline still runs without LLM configuration.

Use autob init to write the global config at ~/.autob/config.json. Project-level .autob.json can override non-sensitive fields such as baseUrl, model, and timeoutMs, but it must not contain apiKey.

Responsibility Boundary

The deterministic engine is responsible for:

• SOP normalization and operation atomization
• field extraction, alias normalization, and knowledge-base enrichment
• operation hypergraph construction
• R1-R10 requirement generation from hypergraph evidence
• fingerprint-based deduplication
• coverage matrix, warnings, Mermaid diagrams, and reports

The LLM layer is responsible only for optional Stage 4 assistance:

• proposing implicit candidate requirements from the existing requirement table
• rewriting candidate text into a concise engineering requirement sentence
• checking likely semantic duplicates before adding a candidate

LLM output is source-bounded. A candidate must cite a source hyperedge, remains candidate by default, and does not replace deterministic confirmed requirements. If the LLM fails, AutoBiology keeps the deterministic output and records a clarification.

GitHub Artifacts

The GitHub repository includes two generated reference directories:

• publication/: paper-oriented methods text, reproducibility notes, sample outputs, Mermaid diagrams, and a draw.io-compatible algorithm figure.
• graphify-out/: generated code knowledge graph, architecture report, and interactive graph HTML.

These directories are source repository artifacts. They are useful for review, reproducibility, and paper writing; the npm package focuses on the CLI runtime, docs, and bundled data/ knowledge base.

Useful Commands

autob config show
autob run your-sop.txt -o out
autob infer out/04-requirements.json -o out

Open out/report.md for a readable summary, then inspect out/04-requirements.json, out/06-clarifications.json, and out/run-meta.json for LLM-specific evidence.