1
I'm a developer and a bit of a knife nerd. I wanted to find out which chef knives are truly the best according to the expert community at r/chefknives. Instead of just counting keywords, I built a multi-pass analysis pipeline to extract brands, models, and steels, and then run sentiment analysis on them.
I analyzed over 1,000 posts from the subreddit. You can see the full results and play with the data here: https://new.knife.day
The Technical Approach: A Hybrid NER Pipeline The core of the project is a 5-phase pipeline that combines fast, deterministic matching with the contextual power of LLMs.
Phase 1: Known Entity Recognition (Fuse.js)
First, I do a quick pass using Fuse.js for fuzzy string matching.
It checks the text against a pre-loaded list of ~465 brands, ~8,700 models, and 50 steel types from an external API.
This is super fast and catches 80-90% of the common entities like "Wüsthof," "Shun," or "VG-10," even with typos.
Phase 2: LLM Entity Discovery (OpenRouter)
To find the niche, artisan, or misspelled brands that Fuse.js misses, I use an LLM.
Crucially, I first mask the entities found in Phase 1 (e.g., "I love my [FOUND_ENTITY] gyuto..."). This forces the LLM to focus only on the unknown terms, saving tokens and preventing redundant work.
I send the masked text to a model like Claude or GPT-4 with a specialized "knife expert" prompt, asking it to extract any remaining brands or models. This is how it discovers less common makers like "Shiro Kamo" or "Yoshikane."
Phase 3 & 4: Sentiment & Summarization (LLM)
With a complete list of entities from both phases, I make another LLM call to perform sentiment analysis on each one, scoring them from -1.0 (negative) to +1.0 (positive).
The system also generates a summary of the entire Reddit thread and calculates a "controversy score" based on sentiment variance.
Phase 5: Storage (MongoDB)
Finally, all the extracted mentions, aggregated entities, sentiment scores, and summaries are saved to MongoDB for analysis.
So, What's the Best Chef Knife According to Reddit? The data reveals some interesting trends:
Most Mentioned: Tojiro, Victorinox, and Takamura are the most frequently discussed brands, often recommended for their value.
Most Loved (Best Positive-to-Negative Ratio): This is a better metric for quality. The clear winners are Japanese artisan brands. Shiro Kamo had a stunning 58 positive mentions for every 1 negative. Fujitora and Masutani also had overwhelmingly positive feedback.
Most Controversial: Shun is the most polarizing brand by far, with a high number of both very positive and very negative reviews. Dalstrong had the worst overall sentiment, with more negative mentions than positive.
One of the biggest challenges was handling the long tail of niche entities. Using Fuse.js alone missed too much, and using an LLM for everything was too slow and expensive. The hybrid two-pass approach provided the best of both worlds.
I wrote a more detailed article about the architecture and findings here: https://new.knife.day/blog/reddit-chef-knives-sentiment-analysis
and more technical details of the analysis can be found here: https://github.com/pvijeh/reddit-named-entity-recognition/blob/main/chefknives-brands.md
Happy to answer any questions. Thanks for checking it out!
The project is a 5-phase system built with Node.js. It first uses Fuse.js for fast, typo-tolerant fuzzy matching of ~450 known brands and ~8,700 models. The remaining text is then passed to an LLM (via OpenRouter) for discovering new, unknown entities and performing sentiment analysis on every mention. I ran it on over 1,000 threads, totaling more than 25,000 comments.
A few interesting findings:
The Underdog: Budget-friendly Tojiro has a massive 27-to-1 positive-to-negative mention ratio.
The Controversy King: Shun is by far the most polarizing brand, sparking strong love/hate discussions (59 positive vs. 24 negative mentions).
The Unloved: Dalstrong was one of the few brands to receive more negative mentions than positive.
The system isn't perfect—I'm open about a critical entity aggregation bug in the write-up. The full technical architecture, results, and raw data are available.
I'm here to answer any questions!
Blog Post (full story & visualizations): https://new.knife.day/blog/we-analyzed-25000-reddit-comments...
GitHub (technical breakdown & raw data): https://github.com/pvijeh/reddit-named-entity-recognition/bl...
Original Reddit Discussion: https://www.reddit.com/r/chefknives/comments/1o2p363/i_analy...