I made a chrome extension that uses Llama 8B and 70B to help avoid BS brands on Amazon
Posted by eclinton@reddit | LocalLLaMA | View on Reddit | 27 comments
I'ts mindblowing how much faster Llama hosted on deepInfra is versus OpenAI models. It takes about 10 seconds to score a new brand. I'm using 8B to parse brands out of product titles when the brand isn't listed on the amazon product, and use 70B for the actual scoring. So far my prompts have performed really well.
The extension has also been surprisingly helpful at exposing me to new quality brands I didn't know about. LMK what you think!
https://chromewebstore.google.com/detail/namebrand-check-for-amazo/jacmhjjebjgliobjggngkmkmckakphel
thehalfblo0dprince@reddit
Is there any source/documentation how to integrate this into chrome extension?
eclinton@reddit (OP)
This is a chrome extension. Not sure I understand the question.
thehalfblo0dprince@reddit
I meant what npm package did you use
eclinton@reddit (OP)
To connect to Llama you mean? I don’t use any package, just the openai-compatible endpoint at deep-infra. Feel free to send me a message directly with what you’re trying to achieve and I’ll try and help.
thehalfblo0dprince@reddit
Got it. So you are indeed using APIs. I was looking for local llm setup like https://www.npmjs.com/package/@mlc-ai/web-llm or https://huggingface.co/docs/transformers.js/en/index
eclinton@reddit (OP)
Understood. I'm considering using these to aid with some minor tasks at the browser level, but for our needs the smaller LLMs just aren't enough. Llama 70B does a lot of the heavy lifting for us.
ArsNeph@reddit
While this is a cool idea in theory, but I have some concerns. Being a brand name does not ensure a quality product, and often comes with insane markups. This is just an example, but look at Louis Vuitton, which sells bags made of Canvas, which is mixed PVC and cotton, one of the cheapest materials to manufacture, and uses cheap, outsourced, but professional stitchwork, then sells the bag that cost $100 to make at most, for $2000. Even though legitimate tailors, leather workers, and the like could make similar or higher quality products for much less, branding allows Louis Vuitton to outcompete these small businesses. Your product would have the side effect of alienating small businesses, OEM extra stock sellers, and startups, further cementing monopolies and established players' positions. Preventing competition is not conducive to better products in any space.
I would say that an extension that can find the model number of the Amazon product, and accurately search the web for specifications, then generate comparison tables with alternative options would allow the user to better make an informed decision. Scraping reviews for sentiment analysis and flaws may also be good. In the case of resellers and fake Chinese brands, it would be good to add functionality that searches the web for that exact product, which is usually found on Alibaba/AliExpress, and inform the user that the product is being resold, and can be bought cheaper directly from the manufacturer.
eclinton@reddit (OP)
Hi, thanks for your comment as it gives me a chance to address an important aspect of the extension. There is little to no value in the extension awarding an “A” to Apple, Nike, Louis Vuitton, Bose, or any other large brand. Those names are easily recognized and you don’t need an extension to tell you who they are. In my usage of the extension I’ve experience the exact opposite effect of what you are suggesting: it has exposed me to many smaller brands that I had not heard about. Our goal is actually to further promote small and medium sized brands, and there’s work we need to do to ensure those are prioritized in our suggestions…. Everyone knows that Nike makes sneakers for kids, but did you know Keen makes amazing ones? Better option than the New Balance we kept buying previously.
The other benefit we’ve seen is when shopping for items in categories I know nothing about. What are good brands of Bike Lights? I can go with a no-name brand that will die in a week, or I can go with NyteRider, a small brand that designs lights that are high quality or durable.
We love small brands and maybe our messaging needs to better reflect that. If you’ve only looked at the screenshots and read the descriptions, I encourage you spend time with the extension. Lots of work left to be done, but perhaps the gap from what you’d find valuable isn’t as wide as you think. Thanks again for the feedback and keep it coming!
ArsNeph@reddit
Firstly, thank you for replying. It's good to hear that you have small businesses in mind, and are optimizing for the best product, over the most brand recognition! Out of curiosity however, I'm wondering how exactly you rank these brands, if not based off of brand recognition and popularity, then what? Are you averaging the ratings of a brand to create a type of ELO system?
eclinton@reddit (OP)
We let the LLM do the heavy lifting based on several prompts, embedding, and tweaks over a few months. Essentially training of what an “A”brand is versus “F” and everything in between. Also gathering data to do fine-tuning. Far from perfect, but the results are good enough to be really useful.
marketflex_za@reddit
Great idea. I'm getting it and look forward to less stressful Amazon browsing.
Can you also make the ads on The Boys go away!??!??!
p.s. I thought of a v2 you could do, too - figure out a way to ignore the products that don't actually sell the item you're looking for but just stuff keywords into it that are related. That has gotten frustrating. New supporter here for sure!
eclinton@reddit (OP)
Haha I can certainly hide ads, I’ve toyed with the idea of hiding sponsored products but don’t want to be too disruptive.
Actually, an interesting goal would be to provide my own results for searches which gives you what you want and eliminates noise. Amazon shows you everything because they make money regardless of what you buy. FakeSpot does this but they show sponsored content versus the best result. Like Bezos, I want to focus solely on the customer and making Amazon more enjoyable to use.
ZebraMoniker12@reddit
very cool, I'm trying it out now. I've used similar extensions like ReviewMeta and FakeSpot in the past to filter out the garbage on amazon and this looks similar
question: as of right now it doesn't tell you the rating until you click the product in the search results and go to the product's page. would it be possible to make it display the rating next to products on the search results page so I can see many at once?
eclinton@reddit (OP)
Yes, that’s the goal! I do this now in certain search result pages when the brand is shown. And it will become more useful as I cache more results wit the product ID (asin). More changes are coming and it will get better as more people use it.
my_name_isnt_clever@reddit
Wow, this is actually so cool. Doing what amazon themselves should have done a decade ago.
first2wood@reddit
I am curious about how this works. It looks great.
eclinton@reddit (OP)
Thanks. We are essentially scraping the brand info from the page, or figuring it out based on the product title… We then have prompts on the LLM for scoring. Llama 70B has been great stock, but we are planning some fine tuning to increase speed and reduce per token costs. We jump around from 8B and 70B depending on prompt complexity, and I can see using bigger models later as well as we expanded.
first2wood@reddit
So you collected some good/bad examples (titles and commodity descriptions) in your prompts as criteria?
eclinton@reddit (OP)
Yes, exactly. Prompts do need more work as it should be returning more D and F scores, but they are much harder to determine. Finetuning should help
first2wood@reddit
Very smart, I was thinking the reviews or web search for reviews which are too easy to fool.
Ulterior-Motive_@reddit
Any chance of getting a Firefox version?
eclinton@reddit (OP)
Yes, that’s planned along with Safari. Just trying to nail down a core set of features before expanding to other browsers. Thanks for the feedback
-illusoryMechanist@reddit
Fakespot also does this
eclinton@reddit (OP)
I'll give you another example, if you search for "durable kids shoes", Fakespot will show "A" to a lot of the resulting products because the reviews seem real, even though few of the brands are actually known for durable shoes. With my extension I found "Keen", a brand I had never heard about but has been amazing as a biking shoe for my kid.
eclinton@reddit (OP)
Fakespot is awesome and I've used it for years, but their focus is on the product's reviews not the brand itself. So even if it's a product comes from a nonsense brand, they'll still score it an A if the reviews seem legit. NamebrandCheck is about elevating companies that actually design and make products, not just resell stuff with a made-up brand attached.
Many_SuchCases@reddit
Getting this: This item is not available Please sign in to view this item, or choose from thousands of other themes and extensions on the Chrome Web Store
eclinton@reddit (OP)
Are you in Europe? That might be the issue. Need to enable that local and republish.