skardi-skills
Agent Skills for working with Skardi — a backend engine to provide SQL pipelines as http endpoints. These skills give your AI coding agent (Claude Code, Cursor, or any Agent Skills-compatible tool) deep knowledge of Skardi's patterns so you don't have to re-explain them each session.
Check out our demo here.
Available skills
| Directory | Skill name | What it covers |
|---|
skills/auto-context/ | auto-context | Turn a folder of documents, a table you already have, or documents still inside a service into governed, searchable context an agent can query. Hybrid search (vector + full-text + RRF) served over HTTP by skardi-server. Three raw-material entries, one flow: a folder; an existing table (SQLite read directly and read-only, any other datastore piped in as NDJSON); or fetch-and-land — list, fetch each body, reconcile, ingest — where your agent writes the per-source fetch code and the skill fixes the process and its acceptance criteria. Storage is a separate choice: defaults to a local SQLite file the skill creates and owns (FTS5 + sqlite-vec vec0 mirrors kept in sync by triggers); point it at Postgres+pgvector, MongoDB, or Lance when the index should live in a database you already own. Handles prereq checks, model resolution, chunking and embedding inline in SQL, ingest, and retrieval end-to-end across candle / gguf / remote_embed. Never creates schema in a datastore you own — prints the SQL and waits — and never writes to a table given as raw material. |
skills/retrieval/ | retrieval | Answer questions from live data through a running skardi-server with the skardi CLI. Discovers sources, named pipelines, and the table schemas the deployment exposes, runs the question through any semantic search surface first (search-hybrid and friends), then writes read-only SQL against exact qualified table names, checks truncation before trusting counts, and reports with the query attached. Consumes whatever the server already has — it builds no index, writes no data, and starts no servers. |
skills/graph-source/ | graph-source | Connect a property graph (a knowledge graph, a GraphRAG corpus — Apache AGE / openCypher) to Skardi and query it through SQL, end to end: provision AGE with a least-privilege reader role, declare the type: graph source and its views in context YAML, triage registration health (healthy / degraded / refused, and what recovery does and does not answer), write correct queries (cypher_query, graph_schema, the JSON getters), and wire Cypher parameters into pipelines. Encodes the traps that bite in production: positional columns binding (same-typed columns declared out of RETURN order swap silently), no predicate pushdown into view Cypher (the bound lives in the view, RowCapExceeded otherwise), wrong-getter silently-NULL columns, the deliberately-absent ->/->> operators, lowercase-only view names, and the one working {params} pipeline spelling. Read-only by backend enforcement; AGE is the backend from Skardi v0.6.0 on (Neo4j/Kuzu are later milestones). |
skills/graph-rag/ | graph-rag | Answer questions whose evidence is in graph relationships: find or verify the named entities, traverse from them through read-only Cypher, and report the bounds and confidence of the edges used. Needs Skardi v0.6.0 or later, like graph-source. |
skills/skardi-query-log/ | skardi-query-log | Stop rewriting SQL you already got working. Every query goes through ask.py, which checks the index first: if this question has been asked, it hands back the statement that worked last time; if not, it files the one that works now. Reads the audit ledger skardi-server --query-audit-db writes (read-only) — seed_index.py picks unfiled queries out of it, read_failures.py reads the failures for what the system cannot answer, and read_log.py reports the overall shape, including whether the pipelines you built are being called. Turns a question into a real pipeline only when it keeps coming back, and then installs it itself: writes the YAML, restarts, checks /health and that /pipelines lists the new name, and rolls the file back out if the restart fails or the server comes back without it. The agent triggers this on its own; a user never asks for a log analysis. |
skills/document-cache/ | document-cache | When your agent is about to read a local PDF, DOCX, XLSX, PPTX, Markdown or text file, check whether Skardi already holds a parsed copy and, if so, read only the table of contents and the sections the question needs. Otherwise the agent reads the file locally and offers to upload it to your personal workspace — it never uploads without a decision recorded for that folder or an answer given this turn, and every failure falls back to the local read. Consent is remembered per folder in ~/.skardi/doc-cache.json. Needs the Skardi MCP tools (find_cached_document, prepare_document_upload, read_document) connected and a Skardi Cloud deployment with the document cache enabled; it does not query the cache with SQL or build any index. |
A running skardi-server is required. Since Skardi's CLI became a thin HTTP client it holds no query engine and no local execution mode, so every path in auto-context starts a server, and retrieval connects to one that is already running. There is no CLI-only mode.
graph-source and graph-rag both require Skardi v0.6.0 or later. v0.5.0 has neither type: graph nor cypher_query, so neither skill can do anything against it. graph-source was verified against 1f2ecae, which v0.6.0 contains.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
