LLM Prompts for Knowledge Graph

Summarize Note

Write ONE short sentence (max 20 words) summarising a technical note. Title: {title} Key terms: {top_terms} Headings: {headings} Reply with the sentence only, no labels.

Rank Knowledge Files

Select the smallest, highest-value set of files needed to understand this project’s knowledge, domain model, architecture, and core behavior.

Goal:

• Keep files that explain WHAT the project does, HOW the main system works, and WHERE the important domain logic lives.
• Prefer files that define architecture, data models, APIs, workflows, algorithms, configuration schemas, or business/domain concepts.
• Select files that would help a new expert understand the project quickly.

Strongly prefer:

• Core source files: .py, .cpp, .hpp, .h, .cc, .cxx, .ts, .js, .go, .rs
• Entry points, main services, routers/controllers, orchestrators, pipelines, engines, managers
• Domain models, schemas, interfaces, abstractions, protocol definitions
• Important config files only if they define project behavior, architecture, or domain concepts
• README or docs only if they contain substantial project-specific architecture/domain knowledge

Deprioritize or exclude:

• Environment/tooling files: .conda, venv, .venv, node_modules, pycache, .idea, .vscode
• Lockfiles and dependency manifests unless uniquely important to understanding architecture
• Build outputs, generated files, caches, logs, binaries, datasets, notebooks with only experiments
• Generic CI/CD, formatting, linting, launch settings, editor settings
• Tests unless they are the clearest documentation of critical behavior or domain rules

Selection rules:

• Return at most {max_files} file indices.
• If many files look relevant, choose the most concept-defining and central ones.
• Prefer a diverse set that covers the main subsystems instead of many near-duplicate files from one area.
• Do not select files merely because they are large; select files because they are semantically important.
• If unsure, favor files closer to core runtime/domain logic over peripheral utilities.

Output requirements:

• Return valid JSON only.
• Do not include explanations, markdown, comments, or trailing text.
• Use exactly this schema: keep
• Indices must be integers from the provided file list.
• Do not invent indices.
• The keep array may be empty if no files are relevant.
• Before selecting each file, mentally ask: would removing this file significantly reduce understanding of the project’s domain or architecture?