gaik-project/gaik-toolkit
GAIK agent skills. For developers building document pipelines with the gaik Python toolkit: choosing a document parser that preserves table structure, designing extraction schemas that stay inside provider limits, building hybrid pgvector search whose failures are visible instead of silent, and writing source-cited reports from mixed files with the gaik report writer. Ready-to-run use cases: turning a short brief into an editable slide deck with speaker notes, timing and citations, and Claude Desktop skills that write incident reports, construction site diaries, priced sales orders and structured reports.
Builds a visual, editable PowerPoint (.pptx) deck with speaker-ready notes, exact timing, citations and a layout-checked design from a topic, an audience and a length, using only the user's own material (strict mode) or adding web research (extended mode). Use when the user asks for a presentation, slide deck, lecture, training session or talk, wants a brief, outline or source documents turned into slides, or needs speaker notes and per-slide timing for a deck. Also use to make targeted changes to a deck this skill produced.
Extracts structured data from Finnish construction site daily diary audio recordings (Työmaapäiväkirja) and creates a formatted Word document with extracted fields. Use when the user needs to process construction site manager's audio diary recordings into standardized documentation.
Extracts structured data — fields, tables, line items — out of documents into a validated schema using the gaik toolkit, and designs schemas that stay inside provider limits and produce checkable evidence. Use when pulling fields from invoices, forms, contracts, reports, or scanned documents into JSON or a Pydantic model; when an extraction request fails with a 400, returns a truncated list, or silently drops its deepest fields; when deciding whether to parse a document first or feed the PDF straight to the model; when adding page numbers, quotes, or confidence scores as evidence; or when extraction accuracy needs to be measured rather than guessed. Covers schema design, property limits, prompt-versus-schema tuning, and citation verification.
Writes a source-cited report from the user's files by running the gaik ReportWriter pipeline: every file (PDF, Word, spreadsheet, text, recording, image) is normalized to text, facts are curated per report section with quotes checked against the sources, and each section is drafted and reviewed. Drafts the report's sections from an optional template or proposes them, and agrees the structure with the user in chat; reports what the sources do not cover, where they disagree, and which facts were dropped. Use when a report must be written from source material and every statement must trace to a source, for example a condition or inspection report, an assessment, or a project or meeting report filled into a fixed template. Also use to correct curated facts and write the report again, or to rebuild a report after its sections were edited. Runs Python with gaik[report-writer]; in Claude Desktop without code execution, use the report-writing skill instead.
Extracts structured data from incident/safety observation audio recordings for manufacturing company reporting. Use when the user needs to process supervisor or employee audio recordings into standardized incident report data.
Converts PDFs, scans, and Word documents into text or markdown with the gaik toolkit's parsers, choosing the parser that will not silently destroy the structure the downstream task depends on. Use when reading a PDF or DOCX into text, pulling tables out of a document, running OCR on scans, feeding documents into a RAG pipeline or an LLM, deciding between PyMuPDF, Docling, and vision-LLM parsing, or when a parse appeared to succeed but the tables, columns, or whole pages came out wrong or empty. Also use when document parsing is costing more time or money than expected. Covers parser selection, per-page verification, and escalation from cheap local parsing to vision models.
Process Purchase Orders and Bills of Material to generate priced Sales Orders
Converts scattered documents and media (recordings, notes, diagrams, PDFs, spreadsheets) into structured MS Word reports. Automatically infers appropriate sections from content or follows provided template/sample structure. Use when the user needs to create a report, summarize documents, consolidate materials, generate documentation, or create meeting notes.
Builds and debugs retrieval with the gaik toolkit — PgVectorStore, Ranker, FinnishTextProcessor, RelevanceGate — as hybrid search: pgvector similarity plus Postgres full-text, fused by rank, and the same design in plain SQL for a TypeScript app. Use whenever gaik is used for search or RAG retrieval, whenever Finnish text is indexed for full-text search, when adding semantic or hybrid search over document chunks, and when deciding whether a search found anything relevant at all. Also use when search misbehaves without an error: a sentence query gets no keyword hits, a word visibly in the text is not found, hybrid returns what vector-only returns, gibberish fills a page of results, or a similarity threshold never filters anything; and when measuring retrieval with Hit@K or MRR, or judging whether a reranker, synonyms, or query translation helped.