The Unstructured Recipe Problem

In the digital age, even personal data can become a chaotic archive. One user faced this reality with a sprawling text file containing years of recipes. This file, originally a simple `.txt`, was a personal data swamp. It housed a mix of complete recipes, ingredient lists typed as single lines, links to external recipes, and even a list of air fryer cooking times. The organizational logic was internal and idiosyncratic, with categories like "Cold," "Slow cooker," and "Complex" that didn't always align with the content. The user's goal was ambitious: transform this digital detritus into a well-structured markdown document with clear headers, improved formatting, and enhanced overall logic.

To achieve this, the user decided to put seven free, web-based chat LLMs to the test. The same source file and identical instructions were provided to each AI. The aim was to see which models could best untangle the mess, reformat the content, and impose a more logical structure on the disparate recipe information. This wasn't just about tidying up; it was a practical experiment in the current capabilities of accessible AI tools for complex, unstructured data transformation.

Testing the Free LLM Landscape

The user's criteria for selecting models were straightforward: they had to be free to use via a web interface. The chosen contenders represented a cross-section of prominent AI offerings:

  • Gemini Flash Extended: This model presented some initial version confusion, identifying itself as 3.5 Flash while stating that 3.1 and 3.5 did not exist.
  • Deepseek Instant w/ Deepthink: The extended version of this model did not support attachments, a limitation for direct file processing. It identified itself as V3 with search restrictions.
  • Claude 3 Haiku: Known for its speed and cost-effectiveness, Haiku was included to see how it handled the unstructured data.
  • Claude 3 Sonnet: A more capable sibling to Haiku, Sonnet was expected to offer a deeper understanding and better output.
  • Claude 3 Opus: The most advanced of the Claude 3 family, Opus was the benchmark for high-end performance in this free tier test.
  • ChatGPT 3.5 Turbo: A widely recognized and accessible model, it served as a baseline for common chatbot performance.
  • ChatGPT 4o: The latest offering from OpenAI, touted for its multimodal capabilities and improved reasoning, was included as a top-tier free option.

The instructions given to each LLM were specific: convert the provided text file into a markdown document. The desired output included a clear hierarchy of headers, improved readability, and a more logical organization of recipes and related information. The user explicitly sought to move beyond the original file's chaotic grouping and formatting.

Performance Analysis: What Worked and What Didn't

The results varied significantly across the seven models, highlighting distinct strengths and weaknesses in handling unstructured, personal data. The core challenge was not just parsing the text but understanding the implicit relationships and the user's intent for reorganization.

Gemini Flash Extended struggled to maintain context and often produced output that was only partially formatted. It frequently lost the distinction between full recipes and mere ingredient lists, and its header structure was inconsistent.

Deepseek Instant, hampered by its inability to process attachments directly in the extended mode, required copy-pasting. While it managed to create markdown, the organization remained rudimentary, often failing to group related recipes or extract information like air fryer times effectively. Its identification as V3 with search restrictions might indicate limitations in its underlying architecture for this task.

The Claude 3 family showed more promise. Haiku, while fast, produced output that was better structured than Gemini or Deepseek but still lacked the nuanced organization the user desired. It often defaulted to a single large markdown file without distinct sections for recipe types.

Claude 3 Sonnet delivered a more coherent structure. It managed to create better headers and differentiate between recipe types more effectively. However, it still occasionally merged distinct recipe entries or failed to accurately extract all the ancillary information, such as the air fryer timings, into a dedicated section.

Claude 3 Opus, as expected, performed the best among the Claude models. It produced the most logical grouping of recipes, created well-defined markdown headers, and made a commendable effort to separate different types of content, including the air fryer times. Its output required the least amount of manual correction, demonstrating superior comprehension of the complex instructions and the messy source material.

ChatGPT 3.5 Turbo offered a mixed bag. It generated markdown with reasonable headers but often failed to capture the nuances of the recipe organization. It tended to oversimplify, treating distinct recipe types as similar and not always correctly identifying shorthand ingredients versus full instructions.

ChatGPT 4o, the most advanced free model tested, demonstrated the strongest grasp of the user's intent. It produced the most accurate and well-organized markdown output, including distinct sections for different recipe categories, well-formatted individual recipes, and a separate, correctly parsed section for air fryer times. Its ability to infer relationships and apply consistent formatting across the entire document was notably superior.

A surprising detail emerged: while many models could convert text to markdown, few could infer the deeper organizational logic the user sought. They excelled at surface-level formatting but struggled with the implicit