Local Knowledge Base Retriever
Finds and cites answers from a local knowledge base through progressive search.
The evidence scopes the skill to local knowledge-base retrieval, asks for confirmation when the default directory is unavailable, emphasizes windowed reads, and requests source attribution. No malware or credential theft is shown. Points are deducted because sensitive-data handling, permission isolation, confirmation boundaries, temporary-file cleanup, rollback, and dependency-integrity checks are unspecified, while PDF/Excel workflows may write files.
The evidence provides a fairly complete workflow with hierarchical navigation, a five-round bound, and explicit handling of insufficient information. Points are deducted for inconsistent tool naming (Read/Grep/Glob versus read_file), unverified tool availability, thin failure handling for PDF/Excel processing and the fallback script, and partially inconsistent retrieval rules. Static review caps this score at 10.
The evidence clearly identifies the audience, paths, default knowledge/ directory, supported file types, trigger wording, and Chinese/English response behavior, and the core workflow does not depend on overseas services. Points are deducted for limited non-fit boundaries, tool-environment assumptions, permission constraints, and large or complex-file edge cases; the semantic trigger remains broad.
The evidence includes layered documentation, bilingual READMEs, references, a script, manifest version 1.0.1, an MIT license, and repository-level validation and release workflows. Points are deducted because the skill does not clearly identify a maintainer, changelog, compatibility verification results, or systematic troubleshooting; version and governance evidence is mainly repository context.
The evidence describes an actionable process for index navigation, progressive retrieval, PDF/Excel preprocessing, bounded iteration, and source-aware answers, which could reduce context usage and produce usable responses. Points are deducted because no representative outputs or behavioral tests are provided, tool portability and complex-format accuracy are unverified, and results still require review. Static review caps this score at 7.
The evidence is auditable through SKILL.md, READMEs, references, the script, manifest, and CI files; CI verifies listing, packaging, and README synchronization. Points are deducted because no committed tests cover the key retrieval paths and no third-party execution evidence is supplied, so actual behavior cannot be confirmed statically. Static review caps this score at 5.
- Confirm that the target agent actually provides the documented Read/Grep/Glob or read_file tools and dependencies such as pdftotext, pdfplumber, and pandas.
- Before using the skill on personal or confidential knowledge bases, add access control, data minimization, temporary-output cleanup, and logging rules.
- PDF/Excel extraction may fail or be incomplete; file locations, inferences, and uncertainty in answers still require human review.
- The mandatory references step is procedural guidance, not a security boundary; it does not provide dependency pinning or protection against malicious content.
What does this skill do, and when should you use it?
kb-retriever helps an agent answer questions using local Markdown, text, PDF, and Excel files. It first follows layered data_structure.md indexes to narrow the search space, then applies file-type-specific retrieval workflows. PDF and Excel processing requires reading the relevant reference guides before extraction or analysis. Retrieval is limited to five iterations and the final answer is expected to identify sources and uncertainty.
It checks the current project's knowledge/ directory by default or uses a path supplied by the user; reads data_structure.md files hierarchically; searches Markdown, text, and log files with grep and local context reads; reads pdf_reading.md before processing PDFs and uses pdftotext, pdfplumber, or related tools for extraction; reads excel_reading.md and excel_analysis.md before processing spreadsheets and uses pandas for exploration, filtering, and analysis; then assembles evidence into an answer with filenames and approximate locations.
- A developer needs an answer from a multi-level local documentation tree covering APIs or system architecture.
- A researcher needs to locate a field, section, or table in local PDF reports.
- An analyst needs to filter or aggregate an Excel report by columns or time range.
- A team needs evidence from a large Markdown or text repository without loading whole files.
- A user needs to search a knowledge base stored at a custom path instead of the default knowledge/ directory.
What are this skill's strengths and limitations?
- Designed for local retrieval without web searching.
- Uses layered data_structure.md files to reduce irrelevant file reads.
- Enforces a learn-before-process workflow for PDFs and Excel files.
- Documents concrete workflows using grep, pdftotext, pdfplumber, and pandas.
- Limits retrieval to five rounds and emphasizes source-aware answers.
- Requires an accessible local knowledge base and cannot safely guess a missing root path.
- Requires shell access, filesystem access, and PDF/Excel processing tools.
- The source provides no test suite, performance metrics, or detailed permission model.
- Its documented workflows cover only the file types explicitly described in the source.
How do you install this skill?
Use the repository's documented skills CLI method: npx skills add ConardLi/garden-skills -s kb-retriever -a claude-code. The README also documents copying the skills/kb-retriever folder into the target agent's skills directory; the exact directory depends on the agent.
How do you use this skill?
In a compatible agent, ask for a knowledge-base retrieval task, such as: “Answer from the knowledge base: what is the API timeout configuration in the system architecture?” The default root is knowledge/ in the current project; provide another path explicitly when needed. If the default directory is missing or inaccessible, the Skill is intended to ask for the correct path.