export_db() keeps documents with no records, as one row with empty record columns. It previously used an inner join, so documents where extraction found nothing were missing from the export (#149).

LLM calls whose structured output cannot be parsed (for example, stray markup after the JSON object) are retried on the same model before falling back to the next one. Content refusals are still not retried.

process_documents(document_id = ...) finds documents from any subdirectory of the project. Stored file paths are relative to the project root but were resolved against the working directory, so reprocessing by ID from elsewhere reported "file not found".

refine_records() no longer reports "Refinement failed: object 'error_log' not found" when every record in a document is human-edited or deleted. The step is now marked skipped.

save_document() now correctly includes id (UUID) when inserting into record_edits. Since 0.1.13 the column is TEXT PRIMARY KEY (NOT NULL), so omitting it caused every cell edit followed by Verify to fail with a NOT NULL constraint error and roll back the entire save (#133).

Extraction write failures are no longer silently swallowed. When save_records_to_db() throws, the error message is now written into extraction_status (e.g. "Extraction failed: ...") rather than leaving the document marked "completed" with 0 records extracted (#134).

migrate_ecoextract_database() now correctly upgrades databases whose records table was created by dbWriteTable rather than init_ecoextract_database(). The previous approach patched the DDL string from sqlite_master, which silently failed when column names were backtick-quoted or AUTOINCREMENT was absent. The migration now builds the new DDL from PRAGMA table_info and uses ALTER TABLE ... RENAME TO / CREATE TABLE / dbAppendTable / DROP TABLE instead of DROP + CREATE + dbWriteTable, so the schema change is guaranteed regardless of how the original table was created.

Records with NULL id (possible when the original table was created without AUTOINCREMENT) now receive a fresh UUID during migration instead of remaining NULL.

Schema migration warning is now issued from get_records() (as a message, printed immediately) and save_document() (as a hard error) rather than from init_ecoextract_database(), which is only called when creating new databases. Old-schema databases opened for reading or writing are now always surfaced correctly.

records.id is now TEXT PRIMARY KEY (UUID v4) instead of INTEGER PRIMARY KEY AUTOINCREMENT. UUIDs are generated in R using base functions only (no new dependency). This makes records collision-free when merging databases from independent researchers.

Added migrate_ecoextract_database() to upgrade existing databases created before 0.1.13. The function recreates records and record_edits with TEXT ids, backfills UUIDs from a per-session integer-to-UUID map, and runs the entire operation in a single transaction. Safe to re-run.

init_ecoextract_database() now warns when it detects the old integer-id schema, prompting the user to run migrate_ecoextract_database().

All record insert paths (bulk extraction via save_records_to_db() and single-row user additions via save_document()) now generate UUIDs for rows that do not already have an id.

Fixed document_id <- NA clobber in ocr_document() that caused reprocessing via document_id to create a duplicate documents row when the file hash had changed (e.g. after opening and re-saving the PDF).

process_documents(document_id = X, force_reprocess_ocr = TRUE) now returns a clear "file not found" status and leaves the database untouched when the stored file path no longer exists, instead of nullifying ocr_status and failing mid-pipeline. A subsequent run without force_reprocess_ocr will use the stored OCR content.

PDF file paths are now stored relative to the project root (found by walking up from the PDF file's directory until a .git, .Rproj, DESCRIPTION, or .here marker is found). This makes paths portable across machines and when the project folder is moved. If no project root marker is found the absolute path is stored as a fallback — this is machine-specific and will break if the database is shared or the project folder is moved.

When a document is identified by hash match but the file has moved to a new path, file_path in the database is immediately updated to reflect the current location.

Extraction now uses a two-turn LLM conversation: turn 1 produces structured reasoning (capturing document analysis, organism identifiability decisions, and extraction choices), turn 2 uses that reasoning as context to extract structured records. This ensures reasoning causally precedes extraction rather than being a sibling field in the same structured output call, fixing divergence between stated reasoning and extracted records.

reasoning removed from the extraction JSON schema — it is now always captured via the two-turn structure and stored in the database regardless of schema configuration.

Content refusals on turn 1 are now correctly detected and logged as refusals rather than retried as stochastic empty-reasoning failures.