require "anthropic" # Reads scanned logbook pages with Claude and returns the entries on them. # One call handles one batch: several page images, or one PDF. # # LogbookTranscriber.new.transcribe([ { data: bytes, media_type: "image/jpeg" }, … ], carried: [...]) # # `carried` are entries already read from page 0 of this batch (the last page of # the previous one), so an entry continuing onto the next page is completed # rather than duplicated. class LogbookTranscriber class Error < StandardError; end MODEL = ENV.fetch("LOGBOOK_AI_MODEL", "claude-opus-5-5") EFFORT = ENV.fetch("LOGBOOK_AI_EFFORT", "high") # handwriting rewards a careful read MAX_TOKENS = 64_000 PROMPT = File.read(File.join(__dir__, "logbook_transcriber/prompt.txt")).freeze SCHEMA = JSON.parse(File.read(File.join(__dir__, "logbook_transcriber/schema.json"))).freeze FALLBACK_BETA = "server-side-fallback-2026-07-01" IMAGE_TYPES = %w[image/jpeg image/png image/gif image/webp].freeze MAX_IMAGE_BYTES = 5.megabytes # the API's per-image limit MAX_PDF_BYTES = 30.megabytes # request limit is 32 MB Result = Struct.new(:entries, :non_logbook_pages, :unparseable_pages, :model, keyword_init: true) class << self attr_writer :client # tests swap in a fake def client = @client ||= Anthropic::Client.new end def initialize(client: self.class.client, model: MODEL, effort: EFFORT) @client = client @model = model @effort = effort end def transcribe(pages, carried: []) stream = @client.beta.messages.stream( model: @model, max_tokens: MAX_TOKENS, system_: PROMPT, output_config: { effort: @effort.to_sym, format_: { type: :json_schema, schema: SCHEMA } }, # If a safety classifier declines, the API retries on its default substitute model. fallbacks: :default, betas: [ FALLBACK_BETA ], messages: [ { role: "user", content: content(pages, carried) } ] ) message = stream.accumulated_message case message.stop_reason when :refusal then raise Error, "Claude declined to read these pages" when :max_tokens then raise Error, "the transcription was too long to finish; try fewer pages per file" end text = message.content.select { _1.type == :text }.map(&:text).join data = JSON.parse(text) Result.new(entries: Array(data["entries"]), non_logbook_pages: Array(data["non_logbook_pages"]), unparseable_pages: Array(data["unparseable_pages"]), model: message.model.to_s) rescue JSON::ParserError raise Error, "Claude's answer wasn't valid JSON" rescue Anthropic::Errors::RateLimitError raise Error, "the Anthropic API is rate limiting requests; try again in a few minutes" rescue Anthropic::Errors::AuthenticationError, Anthropic::Errors::PermissionDeniedError raise Error, "the Anthropic API key was rejected" rescue Anthropic::Errors::APIStatusError => e raise Error, "the Anthropic API returned an error (#{e.status})" rescue Anthropic::Errors::APIConnectionError raise Error, "couldn't reach the Anthropic API" end private def content(pages, carried) blocks = pages.each_with_index.flat_map do |page, i| source = { type: "base64", media_type: page[:media_type], data: Base64.strict_encode64(page[:data]) } if page[:media_type] == "application/pdf" [ { type: "text", text: "This PDF holds the logbook pages; page indices count its pages from 0." }, { type: "document", source: source } ] else [ { type: "text", text: "Page #{i}:" }, { type: "image", source: source } ] end end if carried.any? context = carried.map do |e| "- #{e['date'] || 'undated'} (#{e['log_type']}, signed #{e['signoff_person'] || 'unsigned'}): #{e['text'].to_s.truncate(400)}" end blocks << { type: "text", text: "Entries already read from page 0 (it was the last page of the previous batch):\n#{context.join("\n")}" } end blocks << { type: "text", text: "Extract every logbook entry from these pages." } end end