(+) pdf reader

[scalandr]

universe_kb_add_pdf is the added action for a one-shot PDF ingestion. The action decomposes one PDF into several retrievable knowledge assets. Text chunks go to the text embedding space, images to the vision space, and tables to a dedicated table space. The multimodal vector store backs these with three Chroma collections (_text, _vision, _table). At the workflow layer, the action validates the target universe, accepts either a server-side pdf_path or base64 pdf_content, and forwards the request to POST /kbs/{kb_name}/add_pdf together with optional extraction controls: extract_images, extract_tables, persist_extracted_images, and extracted_image_dir. add_pdf_document() uses PyMuPDF (fitz) to open the PDF, assigns a generated document_id, enriches metadata with source, document type, and extraction timestamp, and then iterates page by page. Text is extracted with page.get_text("text") and ingested through the generic add_document() path as text chunks. Images are discovered with page.get_images(full=True), extracted as raw bytes, and optionally persisted to disk under a deterministic asset directory derived from the KB configuration or the explicit extracted_image_dir. Before ingestion, the code computes rich spatial metadata for each image, including bounding box coordinates, page region, normalized position, surrounding text, detected figure/table anchors, and nearby section context. Tables are handled similarly through page.find_tables() when available: extracted rows are serialized to CSV-like text, augmented with positional metadata, and stored as table modality items.

universe_kb_search now supports three modes: direct_search, rolling_window, and agentic_internal_questions. direct_search is the standard (already available) knowledge base search. The two added modes add relevance control on top of retrieval. rolling_window fetches a larger candidate set, deduplicates it by id, source, or document, and evaluates candidates batch by batch using the agent itself as a strict yes/no relevance judge. agentic_internal_questions first asks the agent to generate a small set of research subquestions, searches the KB for each one, deduplicates the union of results, and again filters them through agent-based relevance checks.