Overview The 2023 Annual Sustainability Report documents measurable year-over-year progress across emissions, water, and renewable energy, while laying out a credible pathway toward net-zero operations by 2030. The company pairs near-term operational wins with a capital-backed decarbonization strategy anchored by interim 2025 milestones. Key Findings and Sections - Scope 1 emissions fell 18% year over year, driven by facility electrification and fleet upgrades replacing fossil-fuel-powered equipment. - Manufacturing water usage declined 12%, reflecting efficiency improvements in production processes. - Renewable energy now supplies 64% of global operations, signaling substantial progress on Scope 2 decarbonization. - A 2030 net-zero roadmap is established, outlining the long-term trajectory for emissions reductions across the enterprise. - Interim 2025 targets are defined to provide accountability checkpoints between current performance and the 2030 goal. - Capital expenditure plans are detailed for solar generation and battery storage, indicating financial commitment to clean energy infrastructure. - The 24-page report integrates operational metrics with forward-looking investment strategy, suggesting alignment between sustainability and business planning. - Fleet and facility upgrades are highlighted as primary near-term levers for direct emissions cuts. Key Takeaways 1. Tangible operational progress is underway: double-digit reductions in Scope 1 emissions and water use demonstrate the strategy is already delivering results. 2. Renewable energy adoption at 64% positions the company ahead of many peers, with solar and battery storage investments set to push that share higher. 3. The combination of 2025 interim targets and a 2030 net-zero roadmap provides clear accountability; stakeholders should track progress against these milestones and monitor capex execution on clean energy projects.
How to Summarize PDFs Offline Without Uploading
Tested prompts for summarize pdf offline locally compared across 5 leading AI models.
BEST BY JUDGE SCORE
GPT-5.4
8/10
If you searched 'summarize pdf offline locally', you already know what you don't want: your document leaving your machine. Maybe it's a legal brief, a medical record, a confidential business report, or internal research. Uploading it to a cloud-based AI service means that content touches someone else's server, and for many people that's a hard no.
The good news is that running a local language model to summarize PDFs is genuinely practical in 2024. Tools like Ollama let you pull open-source models such as Llama 3, Mistral, or Gemma directly onto your laptop or desktop. Combined with a simple script or a local UI like Open WebUI, you can paste or pipe PDF text into the model and get a clean summary with zero network calls.
This page shows you exactly how that workflow runs, what prompt gets the best results, how four different local models compare on the same PDF excerpt, and where the approach breaks down. If you have a Mac with Apple Silicon, a Windows machine with a mid-range GPU, or even a decent CPU-only Linux box, you can do this today.
When to use this
Local PDF summarization is the right call whenever the document contains information you are not authorized or willing to share with a third-party service. It also makes sense when you need repeatable batch processing of many files without per-call API costs, or when your workplace security policy prohibits cloud AI tools outright.
- Summarizing confidential legal contracts or NDA-protected documents before internal review
- Processing patient case notes or medical records that fall under HIPAA or similar privacy regulations
- Batch-summarizing proprietary research papers or internal engineering specs on an air-gapped or restricted corporate network
- Reviewing financial statements or M&A documents where leaking content could constitute a compliance violation
- Academics or journalists working with embargoed or sensitive source materials who cannot risk cloud exposure
When this format breaks down
- Your hardware cannot run a capable model: a 7B parameter model needs at least 8 GB of RAM and runs painfully slow on a CPU-only machine older than 2020. You will wait minutes per page and the output quality will degrade noticeably.
- The PDF is scanned image-only with no embedded text layer. Local LLMs read text, not pixels. Without running a local OCR step first (Tesseract, for example), the model receives an empty string and produces garbage.
- You need to summarize very long PDFs (100-plus pages) in a single pass. Most local models top out at 8K to 32K context tokens. A 150-page technical manual will exceed that limit and require chunking logic you may not be set up to handle.
- The summary needs to be legally defensible or audit-ready. Local model outputs are not traceable or version-controlled by default, and smaller open-source models make more factual errors than frontier cloud models on dense technical or legal text.
The prompt we tested
You are an expert document summarizer running locally on the user's device. Summarize the following PDF content without requiring any external services or uploads. Instructions: Return a structured summary with a 2-3 sentence overview, followed by 5-8 bullet points covering key findings, arguments, or sections, and end with a 'Key Takeaways' list of 3 actionable or memorable points. Keep the total length under 300 words and use plain text formatting only (no markdown tables or images) so it renders cleanly in any offline local AI tool like Ollama, LM Studio, or GPT4All. PDF content: [Extracted text from a 24-page PDF titled 'Annual Sustainability Report 2023'] The company reduced Scope 1 emissions by 18% year over year through facility electrification and fleet upgrades. Water usage in manufacturing dropped 12%, while renewable energy now powers 64% of global operations. The report outlines a 2030 net-zero roadmap with interim 2025 targets and detailed capital expenditure plans for solar and battery storage. Produce the summary now.
Model comparison
| Model | Quality | Cost | Latency | Tokens |
|---|---|---|---|---|
| Claude Opus 4.7 | 7 | $0.0527 | 9.1s | 366 / 629 |
| Claude Haiku 4.5 | 7 | $0.0020 | 4.0s | 256 / 342 |
| GPT-5.4 | 8 | $0.0050 | 6.9s | 225 / 294 |
| Gemini 2.5 Pro | 6 | $0.0017 | 14.2s | 236 / 290 |
| Grok 4.1 Fast Reasoning | 7 | $0.0002 | 5.0s | 372 / 210 |
Outputs
SUMMARY The company achieved significant environmental progress in 2023, reducing Scope 1 emissions by 18% through electrification and fleet modernization while increasing renewable energy adoption to 64% globally. The report presents a comprehensive net-zero strategy with measurable interim milestones and substantial infrastructure investments planned through 2030. KEY FINDINGS • Scope 1 emissions reduced 18% year-over-year via facility electrification and vehicle fleet upgrades • Manufacturing water consumption decreased 12%, demonstrating resource efficiency improvements • Renewable energy now accounts for 64% of global operations, up from prior year baseline • Net-zero target established for 2030 with clearly defined 2025 interim performance milestones • Detailed capital expenditure plan includes solar installations and battery storage infrastructure • Multi-year decarbonization roadmap provides transparency on strategic emissions reduction pathways • Fleet modernization and facility upgrades represent core operational sustainability drivers KEY TAKEAWAYS 1. The company is on track with measurable near-term progress (2023-2025), making the 2030 net-zero goal credible through concrete interim targets and capital allocation rather than aspirational promises alone. 2. Renewable energy dominance (64%) combined with facility electrification signals a systematic shift toward clean operations infrastructure, positioning the company competitively as energy costs and regulations evolve. 3. Water reduction (12%) alongside emissions cuts demonstrates holistic environmental stewardship across multiple resource categories, reducing operational risk and aligning with stakeholder expectations for comprehensive sustainability action.
Overview: The 2023 Annual Sustainability Report highlights measurable environmental progress and a clearer long-term decarbonization strategy. The company reports notable reductions in emissions and water use, increased renewable energy adoption, and a defined 2030 net-zero roadmap supported by near-term targets and planned investments. Key points: - Scope 1 emissions fell 18% year over year, driven primarily by facility electrification and fleet modernization. - Manufacturing water usage decreased by 12%, indicating improved operational efficiency and resource management. - Renewable energy now supplies 64% of global operations, showing substantial progress in energy transition efforts. - The report presents a 2030 net-zero roadmap, giving the company a long-term framework for emissions reduction. - Interim 2025 targets are included, suggesting accountability through shorter-term milestones rather than only distant goals. - Capital expenditure plans are detailed for solar generation and battery storage, signaling that the strategy is backed by planned investment. - Facility electrification appears to be a major lever in the company’s sustainability approach, alongside upgrades to transportation assets. - The overall report emphasizes execution, with operational changes and infrastructure spending aligned to environmental goals. Key Takeaways: - The company is making measurable sustainability progress now, not just setting future goals. - Electrification, renewable energy, and efficiency improvements are the main drivers of impact. - The 2030 net-zero plan is more credible because it includes 2025 targets and concrete solar and storage investments.
OVERVIEW The 2023 Annual Sustainability Report highlights the company's significant progress in key environmental metrics. It details successful reductions in emissions and water consumption while increasing the use of renewable energy across its operations. The document also outlines a strategic roadmap for achieving net-zero emissions by 2030. KEY FINDINGS * Scope 1 emissions were reduced by 18% year over year. * The primary drivers for emission reduction were facility electrification and upgrades to the company's vehicle fleet. * Water usage in manufacturing operations saw a significant drop of 12%. * Renewable energy now powers 64% of the company's global operations. * A formal roadmap has been established to achieve net-zero status by the year 2030. * The roadmap includes specific interim sustainability targets set for 2025. * A detailed capital expenditure plan is in place to fund new solar power and battery storage projects. KEY TAKEAWAYS 1. The company is achieving measurable success, with double-digit percentage drops in both emissions and water use. 2. A clear, long-term strategy is in place, guiding actions and investments toward the 2030 net-zero goal. 3. Significant capital is being committed to renewable energy infrastructure, showing a firm commitment to sustainability targets.
The Annual Sustainability Report 2023 details a company's strong progress toward sustainability goals, with an 18% year-over-year reduction in Scope 1 emissions from facility electrification and fleet upgrades. Manufacturing water usage fell 12%, and renewable energy now powers 64% of global operations. It outlines a clear 2030 net-zero roadmap, including 2025 interim targets and capital plans for solar and battery storage. - Scope 1 emissions cut by 18% through electrification of facilities and upgraded fleets. - Water consumption in manufacturing reduced by 12%. - Renewable energy supplies 64% of global operations. - Comprehensive 2030 net-zero emissions roadmap presented. - Interim targets set for 2025 to track progress. - Detailed capital expenditure plans for solar installations and battery storage. - Emphasis on electrification and renewable transitions across operations. Key Takeaways: - Prioritize facility and fleet electrification for rapid emission cuts. - Scale renewables to over 60% to meet interim targets. - Invest in solar and batteries to achieve net-zero by 2030.
What makes these work
-
01Extract text before prompting
Never paste a raw PDF file path into your local model and expect it to read the file. Use a tool like pdftotext (poppler-utils) or PyMuPDF to extract clean plain text first. Clean extraction removes headers, footers, and page numbers that confuse the model and waste context tokens.
-
02Set explicit output constraints in your prompt
Tell the model exactly what you want: 'Summarize the following in 5 bullet points, each under 20 words, preserving all numbers and named parties.' Vague prompts like 'summarize this' produce inconsistent lengths and bury key facts. Constraints force the model to prioritize rather than pad.
-
03Chunk long documents by section
For PDFs over 20 pages, split the text by logical section (contract clauses, report chapters, paper sections) and summarize each chunk separately, then run a final pass that summarizes the chunk summaries. This keeps each call within the model's reliable context window and produces more accurate results than truncating.
-
04Pick the right model size for your hardware
A 7B model (like Mistral 7B or Llama 3 8B) runs on 8-16 GB RAM and is fast enough for practical use. A 13B model needs 16-24 GB but handles legal and technical language noticeably better. Only go to 70B if you have a GPU with 40-plus GB VRAM; CPU inference at that size is impractically slow.
More example scenarios
#01 · Legal: summarizing a vendor contract for a procurement manager
Input
This Software Services Agreement ('Agreement') is entered into as of January 1, 2025, between Acme Corp and VendorX LLC. The term is 24 months with automatic renewal unless either party provides 90-day written notice. Payment terms are Net-30. Liability is capped at fees paid in the preceding 12 months. Governing law is the State of Delaware.
Expected output
Two-year contract starting January 2025, auto-renews unless cancelled with 90 days notice. Payment due within 30 days of invoice. Vendor liability is capped at the prior 12 months of fees paid. Delaware law governs disputes.
#02 · Medical: condensing a clinical trial abstract for an internal briefing
Input
A randomized controlled trial enrolled 342 adults aged 40-65 with type 2 diabetes. Participants received either 10mg of Drug A or placebo daily for 52 weeks. Primary endpoint was HbA1c reduction at week 26. Drug A reduced HbA1c by 1.4% versus 0.2% for placebo (p<0.001). Adverse events were mild and transient. No serious cardiac events were recorded.
Expected output
Drug A produced a statistically significant 1.4% reduction in HbA1c at 26 weeks compared to 0.2% for placebo. The trial ran 52 weeks in 342 diabetic adults aged 40-65. Side effects were minor with no serious cardiac events reported.
#03 · Finance: summarizing a quarterly earnings report section for an analyst
Input
Revenue for Q3 2024 was $4.2 billion, up 11% year-over-year, driven by a 19% increase in cloud services. Operating income was $680 million, a 7% increase from Q3 2023. Net income declined 3% to $510 million due to a one-time restructuring charge of $95 million. Cash and equivalents stood at $2.1 billion at quarter end.
Expected output
Q3 2024 revenue hit $4.2B, up 11% YoY, with cloud growing 19%. Operating income rose 7% to $680M. Net income dipped 3% to $510M after a $95M restructuring charge. Cash position is $2.1B.
#04 · Academic: condensing a research paper methodology section for a lit review
Input
We employed a mixed-methods design combining a cross-sectional survey of 1,200 undergraduate students across six universities with semi-structured interviews of 40 participants. Survey data were analyzed using multivariate regression controlling for age, GPA, and socioeconomic status. Interview transcripts were coded thematically using NVivo 12. Data collection occurred between March and August 2023.
Expected output
Study used a mixed-methods approach: a 1,200-student survey across six universities plus 40 semi-structured interviews. Regression controlled for age, GPA, and socioeconomic background. Interviews were thematically coded in NVivo 12. Data collected March to August 2023.
#05 · Engineering: summarizing an internal architecture decision record for a new team member
Input
The team evaluated three options for the event streaming layer: Apache Kafka, AWS Kinesis, and RabbitMQ. Kafka was selected based on its throughput capacity exceeding 1 million messages per second, native support for log compaction, and the existing team expertise. Kinesis was ruled out due to vendor lock-in concerns. RabbitMQ lacked the required retention guarantees for audit logging.
Expected output
Kafka chosen as the event streaming layer for its 1M+ messages/sec throughput, log compaction support, and team familiarity. Kinesis rejected over vendor lock-in. RabbitMQ disqualified because it cannot meet audit log retention requirements.
Common mistakes to avoid
-
Feeding a scanned PDF directly
Scanned PDFs contain images, not text. Passing them to a local LLM without OCR preprocessing results in an empty or garbled input. Run Tesseract or a similar local OCR tool first to generate a text layer, then pass that text to the model.
-
Ignoring context window limits
Pasting an entire 50-page PDF into a model with an 8K context window silently truncates everything past the limit. The model summarizes only the first portion without warning you. Always check your extracted text length against the model's context limit before running the prompt.
-
Using a quantized model that is too small
Heavily quantized models (Q2 or Q3 GGUF) that fit in very low RAM produce summaries that drop facts, hallucinate figures, and miss key clauses. For anything accuracy-critical, use at least a Q4 or Q5 quantization. The file size difference is modest but the quality gap is significant.
-
Not verifying numbers and names in output
Local models hallucinate specifics like dollar amounts, percentages, and proper nouns at a higher rate than frontier models. Always spot-check the summary against the source PDF for any numerical claim or named entity before using the output in a professional context.
-
Skipping system prompt configuration
Running the model with no system prompt means it may respond conversationally, add disclaimers, or reformat the output inconsistently across runs. Set a system prompt that defines the role (e.g., 'You are a document summarization assistant. Output only the summary, no preamble.') for consistent, clean results.
Related queries
Frequently asked questions
What is the best local AI model for summarizing PDFs?
For most users, Llama 3 8B or Mistral 7B running through Ollama is the best starting point. Both balance speed, memory requirements, and summary quality well. If you have 24 GB or more of RAM and need better handling of legal or technical language, Llama 3 70B (quantized) is a meaningful step up.
How do I run a local LLM on my laptop without a GPU?
Ollama supports CPU-only inference on Mac, Windows, and Linux. Download Ollama, run 'ollama pull mistral' in your terminal, then 'ollama run mistral' to start a chat session. CPU inference on a modern laptop is usable for short documents but expect 1-3 minutes per page for a 7B model on older hardware.
Does summarizing locally mean the data never leaves my computer?
Yes, if you are running a model locally through Ollama or LM Studio with no cloud sync enabled. The inference runs entirely on your CPU or GPU, no network call is made, and no data is sent to any external server. Confirm your chosen tool has no telemetry or cloud backup feature enabled by default.
Can I summarize a PDF offline on a Mac with Apple Silicon?
Yes, and Apple Silicon is actually one of the best platforms for this. Ollama is native on M1, M2, and M3 Macs and uses the unified memory architecture efficiently. A 7B model runs quickly enough for practical document work, and 13B models are usable on machines with 16 GB or more of unified memory.
How do I convert a PDF to text before summarizing it locally?
On Mac or Linux, install poppler-utils and run 'pdftotext yourfile.pdf output.txt' in the terminal. On Windows, you can use the same tool via Chocolatey or use Python with the PyMuPDF library (fitz). Both give you a plain text file you can paste or pipe directly into your local model's prompt.
What if my PDF is too long for the model's context window?
Split the PDF by section or by page count into chunks that fit within the model's context limit (typically 4K-32K tokens depending on the model). Summarize each chunk separately, then pass all the chunk summaries into a final prompt asking the model to produce a combined summary. This recursive approach preserves accuracy across long documents.
Try it with a real tool
Run this prompt in one of these tools. Affiliate links help keep Gridlyx free.