// Helper function to process a single PDF const processPdf = wrapTraced( async (pdfFile: { filename: string; url: string }) => { console.log(`Processing ${pdfFile.filename}...`); // Fetch and encode the PDF file from URL (with error handling) let pdfData: Buffer; let base64String: string; try { const response = await fetch(pdfFile.url); if (!response.ok) { throw new Error(`HTTP ${response.status} ${response.statusText}`); } const arrayBuffer = await response.arrayBuffer(); pdfData = Buffer.from(arrayBuffer); base64String = pdfData.toString("base64"); } catch (err: any) { console.error(`Failed to download ${pdfFile.url}:`, err); return; // Skip this PDF and continue with the next } const userPrompt = "Please analyze this earnings call transcript"; const rootSpan = currentSpan(); rootSpan.setAttributes({ name: pdfFile.filename }); const rootSpanSlug = currentSpan().export(); // Create chat completion with file data const completion = await client.chat.completions.create({ model: "gpt-4o", messages: [ { role: "system", content: SYSTEM_PROMPT, }, { role: "user", content: [ { type: "file", file: { filename: pdfFile.filename, file_data: `data:application/pdf;base64,${base64String}`, }, }, { type: "text", text: userPrompt, }, ], }, ], max_tokens: 500, }); const summary = completion.choices[0]?.message?.content; // if no summary is generated, log an error and return if (!summary) { console.warn("No summary generated"); return; } // Console log that the summary was created console.log( `\nEarnings Summary for ${pdfFile.filename}: Summary Created! View in the Braintrust UI!\n`, ); // log the output of the LLM call to the root span rootSpan.log({ output: summary, }); // Log system prompt span await logSystemPrompt(pdfFile.filename, pdfFile.url, summary, rootSpanSlug); // Log user prompt span await logUserPrompt( pdfFile.filename, pdfFile.url, userPrompt, summary, rootSpanSlug, base64String, ); }, logger, );
This method streamlines the process when you’re working with publicly available PDFs, like the earnings call transcripts we’re using in this cookbook. Both methods result in the PDF being attached to your prompt, allowing the LLM to analyze its contents. Choose the approach that best fits your workflow based on where your files are stored.