
339 stars, and not one real person speaking: a job-hunting project's "star-farming + a good README"
What it is A Python project that runs RAG Q&A over vite's GitHub issues and writes the evaluation up as a paper. The author states in the repo that it's a portfolio piece for Chinese-speaking interviewers.
Why it blew up The stars are mostly from a botnet starring at a steady daily pace (6 forks, all bot accounts, and the same accounts are starring 20 to 30 small Chinese repos). The growth curve then got picked up by AI news sites and repost accounts that select repos by growth rate, which made it look like a hit.
CHow you could approach it Don't copy the topic, and don't copy the stars. Take the README approach (counterintuitive numbers, reported ranges, per-item results you can check, stated limitations) and apply it to a problem nobody has tested yet, such as the Chinese retrieval the project admits it can't do. Then get your first real stars from Show HN, V2EX, and X.
- Data (updated daily) · Repo created 2026-11-01 · 482 stars · 7 days +153
GitHub · stars Data 2026-11-06 → 2026-11-20 (15 days, 15 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
A GitHub account registered in Nov 1 got 339 stars in 12 days, and several AI news sites wrote it up as "GitHub trending." Looking closer, those stars can almost certainly be bought. The author was after a glance from an interviewer. The README is genuinely well written, though. Below: where the stars came from, which parts are worth learning, and whether copying this kind of project to get stars is worth it.
A repo created by an account only 4 minutes old
On Nov 1 11:28, 2026 (UTC), a new GitHub account called zaxiu88888 appeared. Its bio was just one line: "Maybe I'm still human after all." 4 minutes later it created the issue-rag repo. All 33 commits are by "ccb," and the earliest is from Oct 27, meaning the work was finished locally before it was pushed. Nothing in the code has changed since Nov 4.
The stars never stopped, though: 11-06 48 → 11-07 101 → 11-08 130 → 11-09 146 → 11-10 188 → 11-11 233 → 11-12 285 → 11-13 329, and by that day the GitHub page showed 339. It gained 281 in one week, with a biggest single day of +53. There was no sudden spike, just forty to fifty added steadily every day. At the same time: 0 issues, 0 watchers, and the author had 0 followers.
What it does: a README written up like an evaluation paper
The topic isn't new: RAG Q&A over vitejs/vite issues. It uses real issue titles as questions and the maintainer's reply when closing the issue as the reference answer. The author hand-checked 95 items and compared five approaches: closed-book, BM25, vector, hybrid, and hybrid with reranking.
What's different is the writing. The first screen of the README is a three-row table:
| No retrieval | Retrieval (best config) | |
|---|---|---|
| Judged correct | 8.4% | 64.2% |
| Weighted accuracy | 0.221 | 0.700 |
| Judged as fabricated | 1.1% | 20.0% |
The most striking line is the last one: adding retrieval actually raised the fabrication rate from 1.1% to 20%, so the model went from "honestly says it doesn't know" to "confidently answers wrong." It also avoids reporting one flattering multiplier and gives a range instead: 2.3× with a stricter judge model, 3.2× with the original judge, 30× on 5.2 blind human-reviewed items. "So the machine-measured number is a lower bound. It is not a 4× result." It also audits the judge itself in four ways: swapping models, running the same model twice, changing the scoring rubric, and blind human review (70% agreement, κ=0.816). It states its own limitations too: retrieval only supports English, so Chinese questions get refused. Finally, there's a GitHub Pages results page where all 95 items can be searched and filtered one by one.
None of these methods are new in the evaluation field, but few individual projects write them up as a complete dataset-backed document that anyone can verify item by item.
Who it's written for: the author says so
This one isn't a guess. A commit message reads: "The repo is going to be read by Chinese-speaking interviewers… the first thirty seconds decide whether anyone reads the measurement at all." (The first 30 seconds decide whether anyone reads the measurement at all.) Another commit adds a Chinese NOTES.zh-CN.md titled 面试速记 ("Interview cheat sheet"), and its first line is:
The typical way AI-assisted projects die: the code runs, but you can't explain it, and the interviewer spots that in 90 seconds.
Below: the "30-second pitch for this project," plus "the trade-offs you have to be able to defend" (why hybrid retrieval, why RRF...). There's also a commit that adds Chinese comments to "the six functions an interviewer is most likely to ask about." The same account created two more repos on Nov 4 and 9: one is an agent platform built with Spring AI, the other a problem-solving notes community built with Spring Boot + React. Together they make up a typical Chinese job-hunting resume project set.
So this project's real "user" is the hiring side, and the 339 stars are there for the resume. That's an inference, but the evidence in the repo is direct.
Where the 339 stars came from
GitHub doesn't publicly show who starred a repo, so you can't check them one by one. But forks are public, and the 6 forks look like they came from one template:
| Fork accounts | Fork date | Account registered | Bio | ||
|---|---|---|---|---|---|
| gritbugaboo8m | 11-05 | 2026-10-05 | Akua Mthembu / Software Engineer — running, coffee | ||
| outlets-flops-56 | 11-05 | 2026-08-27 | Jürgen Bertrand / DevOps Engineer. Love reading & astronomy. | ||
| swath-bastion-8w | 11-06 | 2026-08-08 | Luna Hill / running lover. kayaking… | ||
| bronzer-cornel1f | 11-08 | 2026-10-22 | Drew Allen / Software Engineer \ | coffee \ | hiking |
| dragmailer1u | 11-09 | 2026-08-08 | Liam Anderson / cycling and astronomy. Architect. | ||
| reveriemasted9f | 11-10 | 2026-09-24 | Sandrine Lefebvre / gardening · coffee · Nurse |
The usernames are two English words plus two random characters, paired with a Western first name and stock hobbies. They have 0 followers, every public repo is a fork, and their public activity is nearly all starring and forking, mostly small Chinese-authored repos. About one fork a day, spread evenly, so the fork-to-star ratio looks normal (6 to 339, under 2%).
Following the repos these 6 accounts starred, 24 repos were starred by more than 4 of these accounts at once. Here are a few that are growing:
| Library | Repo created | Daily stars after the rise began | Single-day peak |
|---|---|---|---|
| zaxiu88888/issue-rag | 2026-11-01 | 48 → 101 → … → 329 | 53 |
| TngSHoO/agentcase-lab | 2026-11-04 | 28 → 41 → 99 → … → 344 | 58 |
| qiaoranlaodou-cpu/your-next-point | 2026-11-06 | 97 → 127 → … → 389 | 71 |
| BDiff/BDiff | 2025-09-30 | 162 flat for several weeks → 193 → … → 454 | 63 |
| madong-code/madong-admin | 2025-07-23 | 194 flat → 243 → … → 685 | 72 |
| AnomalyNTE/Anomaly | 2026-07-29 | 95 → 138 → … → 570 | 64 |
Some are brand-new repos, others have sat flat for weeks or are over a year old, but once they start rising they follow the same line: around three hundred a week, forty to seventy a day. issue-rag can't be told apart from them.
One more detail: around November 6 and 7, this group of repos lost stars together. zcbacxc/mcd-party-planner alone dropped 211 in one day (1593 → 1387). That looks like GitHub cleaning out fake accounts. The conclusion that the stars were bought comes from the fork accounts, the curve shapes, and the absence of any real people. I haven't verified each account individually, but that all 6 forks come from bot accounts is certain.
How the "coverage" came about
Searching its name turns up several write-ups that look like a hit. In date order:
| Timeline | Who wrote it | Stars at the time |
|---|---|---|
| From ¥11-05 | Bot accounts start forking | Under 48 |
| 11-07 | amazingindex's "Xiangxu Institute" daily, recommendation score 68 | 102 |
| 11-07 | reveneau.com English news, page says "This item was written by an AI system" | 102 |
| 11-08 | Templated "🚀 Trending on GitHub … ⭐121" posts on Threads and LinkedIn | 121 |
| 11-10 | Listed by "AI Flag" | — |
| 11-11 | A long Chinese thread from @xqliu on X | 246 |
| Always there | The RAG topic chart on trendshift | 339 |
Every one of them came after the stars started climbing. These sites and accounts pick repos automatically by "new + growing fast," the growth rate triggered them, and they then nearly copied the README word for word. The reveneau piece says outright: "The author's own README is the source for every number below."
The only one that reads like a real person wrote it is @xqliu's Nov 11 post: "Someone built a Q&A assistant over vite's GitHub issues…" with 584 views, 3 likes, and 0 reposts. This account has posted an "someone built xxx" introduction of a new repo almost every day for the past two weeks, usually plugging its own product too. After that post, issue-rag still gained +45, +52, +44 a day, with no inflection in the curve. Hacker News, Reddit, and Product Hunt have nothing on it.
That's the second benefit of buying stars: you pay for a growth curve and get back a stack of "as featured in" screenshots.
The same topic, and how many stars others got
Products for "ask questions about an open-source repo's issues" have been mature for a while. Dosu is a GitHub App that installs into a repo and answers issues automatically. Kapa.ai is offered free to all CNCF-hosted projects. Cognition's DeepWiki launched in April 2025 and indexed over 30k repos right away, and its open-source clone deepwiki-open has 18k stars. On the RAG evaluation side, the big frameworks own the space:
| Project | Launched | star | How it differs from issue-rag |
|---|---|---|---|
| promptfoo | 2023-04 | 27k | Testing and red-teaming for prompts, RAG, and agents |
| DeepEval | 2023-08 | 19.5k, +146 a week | General evaluation framework with built-in LLM judging |
| ragas | 2023-05 | 16.3k, +58 a week | General RAG evaluation metrics |
| RAGChecker(Amazon) | 2024-06 | 1.1K | Fine-grained diagnostics that accompany a paper |
| vectara/open-rag-eval | 2024-11 | 423 | Evaluation that doesn't require reference answers |
| knowflow-ai/KnowEval | 2025-11 | 53 | End-to-end evaluation of Chinese enterprise documents, also a small methodology-focused repo |
The last row matters most. KnowEval is also a small evaluation repo focused on methodology. It's been around for a year and has 53 stars (on GitHub, 11-12). Larger frameworks with much more solid methods gain 58 to 146 stars a week. A personal evaluation assignment that took 12 days got 339 stars and not one real discussion, which doesn't match the content.
What to learn, and what not to
What's worth learning is the README's writing:
- Put one counterintuitive number on the first screen (fabrication rate 1.1% → 20%), not a feature list;
- Report a range, not the flattering single number, and say why;
- Ship a static results page people can check line by line, so they can verify it themselves;
- State the limitations up front, in both Chinese and English.
This framing is easy to pass along secondhand: AI news sites and that tweet are copying its numbers directly.
What you can't copy is the stars:
- Bought stars get purged. In one batch, a repo already lost 211 in a single day.
- Any interviewer who clicks through sees 0 issues, 0 watchers, an author with 0 followers, fork accounts that are all fake, and a star count that keeps rising at a steady clip after the code has stopped changing. Anyone in the know spots it immediately, and that's worse than having no stars at all.
- The same boost network is simultaneously packaging twenty or thirty small Chinese repos. This is not some proprietary tactic.
If it were you
Bottom line: you can learn from it, but copying it isn't worth the effort. This project's stars didn't come from the content. If you copy the same topic, you'll most likely end up with a few dozen stars like KnowEval got. And the topic is crowded on both ends: for asking questions about issues there's DeepWiki and Dosu, and for evaluation there's ragas and DeepEval.
Who it's for: developers preparing for AI application and RAG job interviews who want a portfolio project that's clear to explain and holds up under follow-up questions. Using its README structure and its "interview cheat sheet" as a template is well worth it.
Who it's not for: anyone hoping a single open-source project will earn real stars and followers. Copying the topic won't get you stars; copying the boost will get the stars wiped and your reputation with them.
How to do it: don't build the Nth "vite issue Q&A." Apply the same approach to a question nobody has answered with data yet: something people argue about but nobody has rigorously tested, like the real-world behavior of a popular framework's new release, or Chinese-language scenarios. The project itself admits Chinese retrieval is a weakness, which is a ready-made gap. Deliver the same way: one counterintuitive number, a range, a results page you can check line by line, and clearly stated limitations.
How to get it seen: start with real human channels, like Show HN on Hacker News, V2EX, and X, and lead with that one counterintuitive number. Once a few days of genuine stars come in, the aggregator sites and trending lists that pick repos by growth speed will likely find it on their own (inference).
Timing: there's no hard window. If it's for job hunting, aim to finish before the spring hiring season in 2027 (inference).
What would show you misjudged this: if issue-rag keeps gaining stars after its curve stops rising steadily, with real people opening issues, sending PRs, or discussing it on HN or V2EX; or if a small methodology repo that wasn't boosted starts gaining over a hundred stars a week from content alone. That would mean a good README can earn stars by itself, which would be worth following closely.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one Dosu Other 2023 A commercial GitHub App, used by large projects like Apache Superset (per the official case studies) Actually installed in repos to automatically triage and answer issues; issue-rag is just an offline experiment Kapa.ai Other 2023 Offered free to all hosted CNCF projects Connects docs, issues, and forums together for technical support Q&A; issue-rag tests only one repo DeepWiki(Cognition) Big-company features 2025-04 Indexed over 30k repos from the start; the open-source clone deepwiki-open has 18.4k stars You can ask about the code and docs of any public repo, free, at product quality HF Cookbook: simple RAG over GitHub issues Other 2024 Official HF tutorial A starter example on the same topic, with no evaluation ragas Open source 2023-05 16.3k stars, +58 over 7 days A general RAG evaluation metrics framework; issue-rag is one specific experiment plus a write-up of the approach DeepEval Open source 2023-08 19.5k stars, +146 over 7 days A general evaluation framework with built-in LLM-as-judge promptfoo Open source 2023-04 27.1k star A testing and red-teaming tool for prompts, RAG, and agents RAGChecker(Amazon) Open source 2024-06 1.1k star A fine-grained RAG diagnostic framework from a paper vectara/open-rag-eval Open source 2024-11 423 star RAG evaluation that doesn't need reference answers knowflow-ai/KnowEval Open source 2025-11 53 star A methodology-style small repo of the same kind (end-to-end evaluation of Chinese enterprise documents) has only a sixth of its stars, which shows issue-rag's star count didn't come from its content - Where users come from
- Inferred star-farming / fork-farming service: 6 fork accounts (gritbugaboo8m, outlets-flops-56, swath-bastion-8w, bronzer-cornel1f, dragmailer1u, reveriemasted9f) all have "word+word+random characters" usernames, Western-style names, templated bios, 0 followers, and their repos are all forks, with one fork added per day at a steady rate. The same accounts starred twenty or thirty other small Chinese repos, and those curves during the boost period look just like issue-rag's. The stargazer list is hidden by GitHub, so individual stars can't be checked one by one.
- Verified README written to get picked up (counterintuitive number as hook, bilingual, online results page): reveneau writes "The author's own README is the source for every number below"; @xqliu's tweet is nearly a paraphrase of the README's Chinese summary.
- Verified Auto-aggregator sites and trending lists that pick repos by growth speed: the amazingindex daily digest 11-07 (102 stars), reveneau 11-07 (AI-written), Threads/LinkedIn "Trending on GitHub ⭐121" template posts 11-08, aiqipage 11-10, and the trendshift RAG list. All of them came after the star growth.
- Verified X Chinese KOL: @xqliu 11-11 When the tweet went out the repo already had 246 stars; the tweet got ❤3 🔁0 👁584; after the tweet the curve showed no inflection (+45/+52/+44); this account posts about someone else's new repo nearly every day and mentions its own product Orbi along the way
Why issue-rag (zaxiu88888/issue-rag) was able to run this way
Channel: GitHub; research date 2026-11-13 (data card refresh date). Data comes from our github / x / hackernews databases, live GitHub API lookups, x-radar and live X lookups, and web search (Jina / Exa).
Bottom line
This is a portfolio project for job hunting: the author states in the repo that it's meant for "Chinese interviewers" and attaches an "interview cheat sheet." Its star curve and sources can essentially be judged as bought:
- All 6 forks come from the same network of boost accounts;
- Those accounts are also boosting stars and forks for twenty or thirty small Chinese repos, and those repos' curves look exactly like this one (about +300 per week, steadily +40~60 per day);
- Every bit of "spread" you can find online (AI-written news sites, "Trending on GitHub" template accounts, the trendshift list, and one tweet from @xqliu with 584 views) came in response to the star growth speed. All of it came after the stars started rising, and none of it drove the curve.
The content itself is solid: the README reads like an evaluation paper, opens with a counterintuitive number (adding retrieval took the fabrication rate from 1.1% to 20%), and has an online results page you can go through line by line. So when the stars arrived, the auto-aggregator sites were happy to pick it up. It isn't a product that took off on its own. It's a resume package built on star-boosting plus a good README.
1. Basic data
- Created 2026-11-01 11:32; the author's account zaxiu88888 was registered only 4 minutes before the repo was created (GitHub API: account created_at 2026-11-01T11:28). The account has 0 followers and 0 following, with the bio "Maybe I'm still human after all."
- Stars (data card and our github database): 11-06 48 → 11-07 101 (+53) → 11-08 130 → 11-09 146 → 11-10 188 → 11-11 233 → 11-12 285 → 11-13 329; live GitHub API lookup 339. Over the last 7 days +281, with a single-day peak of +53, and no single-day spike.
- Forks 6, watchers (subscribers) 0, open issues 0, no Releases. The last push was 11-04, and the code hasn't changed since, yet stars keep rising at a steady clip every day.
- 33 commits, all under the author name "ccb"; the earliest is 10-27 (built locally first, pushed up on 11-01). The commit messages are very long and tidy, clearly in an AI-written style.
- The same author's two other repos: spring-agentforge (created 11-04, an agent platform built on Spring AI, 1 stars) and notehub (created 11-09, a LeetCode-style note-taking community on Spring Boot + React, 0 stars). Both are classic Chinese job-application resume projects.
- Our github database flagged it with an empty suspect field (no automatic rule was triggered), and its status is "newly appeared." It hasn't appeared on trending (repos under 60 days old can't make it there anyway).
2. What it is: a project written for interviewers
- Content: RAG Q&A over issues in vitejs/vite. It uses real issue titles as questions and the maintainer's reply when closing the issue as the reference answer. 95 items were manually reviewed, comparing five setups: closed-book, BM25, dense, hybrid, and hybrid+rerank. It then audits the grader itself from four angles: swapping models, running the same model twice, changing the rubric, and 30 blind manual ratings (κ=0.816). The final output is a gain range (2.3×–3.2×), not a single flattering number. There is also a GitHub Pages results page you can search and filter. (README)
- The author states the purpose directly:
- Commit "Add a Chinese README next to the English one": "The repo is going to be read by Chinese-speaking interviewers… the first thirty seconds decide whether anyone reads the measurement at all."
- Commit "Add Chinese interview notes": "NOTES.md is the document you rehearse from before an interview." The heading of NOTES.zh-CN.md is "interview cheat sheet," and its first sentence is "Projects built with AI assistance have a typical way of dying: the code runs, but you can't explain it, and the interviewer spots it in 90 seconds."
- Commit "Clean up dead front-end code and add Chinese comments": added Chinese first-line comments to "the six functions interviewers are most likely to ask about."
- In other words, this project's "user" is the hiring side, and the stars are a signal on a resume. The author is trying to win an offer, not traffic or revenue (this is inference, but the evidence in the repo is direct).
3. Where the stars came from: the evidence points to a boost network
3.1 6 fork accounts are all bots
The 6 forks listed by the GitHub API (timestamps in UTC):
| Fork accounts | Fork date | Account registered | Pattern | ||
|---|---|---|---|---|---|
| gritbugaboo8m | 11-05 11:03 | 2026-10-05 | "Akua Mthembu / Software Engineer — running, coffee" | ||
| outlets-flops-56 | 11-05 15:36 | 2026-08-27 | "Jürgen Bertrand / DevOps Engineer. Love reading & astronomy." | ||
| swath-bastion-8w | 11-06 18:16 | 2026-08-08 | "Luna Hill / running lover. kayaking…" | ||
| bronzer-cornel1f | 11-08 14:15 | 2026-10-22 | "Drew Allen / Software Engineer \ | coffee \ | hiking" |
| dragmailer1u | 11-09 13:43 | 2026-08-08 | "Liam Anderson / cycling and astronomy. Architect." | ||
| reveriemasted9f | 11-10 13:03 | 2026-09-24 | "Sandrine Lefebvre / gardening · coffee · Nurse" |
What they have in common:
- Usernames are all "two English words joined + 2 random characters";
- each has a Western first name and a templated hobby bio;
- each has 0 followers and 0 following, and their public repos are all forks;
- their public activity is almost entirely WatchEvent (starring) and ForkEvent, targeting small repos by Chinese authors (BDiff/BDiff, TngSHoO/agentcase-lab, madong-code/madong-admin, qiaoranlaodou-cpu/your-next-point...);
- forks arrive one per day, spread out evenly.
3.2 other "customers" on the same network, with curves cut from the same mold
Of the repos starred publicly by these 6 accounts, 24 were starred by more than 4 of them at the same time. Taking a few of those that are being boosted and comparing them with issue-rag (our GitHub repo, daily star counts since 11-01):
| Library | Repo created | Daily stars | gain_7d | Single-day peak |
|---|---|---|---|---|
| zaxiu88888/issue-rag | 11-01 | 48,101,130,146,188,233,285,329 | 281 | 53 |
| TngSHoO/agentcase-lab | 11-04 | 28,41,99,142,192,214,269,305,344 | 303 | 58 |
| qiaoranlaodou-cpu/your-next-point | 11-06 | 97,127,179,250,296,366,389 | — | 71 |
| BDiff/BDiff (a 2025-year-old repo) | 2025-09-30 | 162 flat for weeks → 193,219,242,305,345,403,454 | 292 | 63 |
| madong-code/madong-admin (an old repo) | 2025-07-23 | 194 flat → 243,266,330,361,406,456,490,562,589,637,685 | 324 | 72 |
| AnomalyNTE/Anomaly | 2026-07-29 | 94,95 → 138,181,245,302,330,387,406,462,514,544,570 | 268 | 64 |
| vladimirrott/mojo-dllm | 2026-10-04 | 244,280,295,339,402,436,469,504,514,520,526 | 187 | 78 |
Some of these repos are new, others have existed for a year, but once the boosting starts it's +300 or so per week, steady at +40~70 per day, and most authors have only this one repo. The curve for issue-rag can't be told apart from theirs.
- This batch of repos collectively dropped slightly in stars around 11-06/07. zcbacxc/mcd-party-planner lost 211 in a single day (1593→1387), and the others each lost 4–10. This matches the pattern of GitHub purging fake accounts, which suggests the network has already been noticed by the platform; issue-rag also risks losing a chunk of its stars later.
- Limitation: GitHub now hides the stargazer list for all repos (REST returns 404, GraphQL returns empty. Trying 0xsline/OpenChatCut gave the same result), so I couldn't check individually who starred the 339 stars. The public starred lists of these 6 accounts also don't include issue-rag, which means the accounts that starred it are other accounts in this network. "The stars were bought" is a strong inference, not something verified case by case. But the fact that 6/6 of the forks are bots is confirmed.
3.3 "spread" all followed the rise in star velocity, and every one lags in time.
| Timeline | Event | Stars at the time | Whether it converts into traffic |
|---|---|---|---|
| From ¥11-05 | Bot forks begin (the accounts were created only two days earlier) | <48 | — |
| 11-06 22:52 UTC | amazingindex "向上研究所" daily digest No.003, "recommendation index 68" (auto-picked + AI commentary) | 102 | Automatic aggregator site, picks repos by star velocity |
| 11-07 00:40 UTC | reveneau.com English news; the page itself says "This item was written by an AI system" | 102 | Same as above |
| 11-08 | Threads / LinkedIn account Arthi Rajendran posting with a "🚀 Trending on GitHub … ⭐121" template | 121 | Templated repost accounts |
| 11-10 | Indexed by aiqipage "AI旗页" | — | Automatically indexed |
| 11-11 10:45 Beijing time | @xqliu (Larry 🍀 Orbi), a long Chinese tweet: "Someone built a Q&A assistant on vite's GitHub issues…" | 246 | ❤3 🔁0 💬5 👁584 (checked on X now) |
| Ongoing | Listed on trendshift's RAG topic chart (339 stars) | — | A chart sorted by star velocity |
- @xqliu is not the promoter of this project. In the past two weeks he has posted almost daily "someone built xxx" introductions to new repos (PyLings, lot-map, Lightpanda, MACaQ...), mostly with 300–600 views and 0–4 likes, aimed at giving his own product Orbi a boost (the second reply to this tweet lands on "Orbi's release version number is decided by humans"). Since he posted, issue-rag has still gained +45 / +52 / +44 per day, and the curve shows no inflection point at all. The data card's claim that "a KOL is running paid promotion on X with a long Chinese post" does not match the posting time, the view counts, or the curve.
- HN (our repo + Algolia), Reddit, and Product Hunt have none of it. None of the posts in the x-radar local corpus mention it either.
- Conclusion: the stars came first, the aggregators came after. Aggregators, repost accounts, and charts all pick products automatically based on "new repo + velocity." Boosting stars triggers them, and they in turn give the project a layer of "it's been covered" appearance. This is a common secondary payoff of star boosting.
4. Competitors: who else does the same thing
"AI Q&A over open-source repo issues" has long been a mature feature, and "RAG evaluation / LLM judging" has plenty of large projects. issue-rag has no product competitiveness in either area; its positioning is "an evaluation assignment."
| Name | Form | Launched | Scale | How issue-rag differs |
|---|---|---|---|---|
| Dosu(dosu.dev) | Commercial SaaS, GitHub App | 2023 | Used by large projects such as Apache Superset (case studies on their website) | Actually installed into the repo, auto-triages and answers issues; issue-rag is just an offline experiment and doesn't connect as a GitHub App |
| Kapa.ai | Commercial SaaS | 2023 | Offered free by CNCF to all hosted projects (kubernetes/website discussion thread) | Connects docs, issues, and forums together for technical support Q&A; issue-rag tests only one repo |
| DeepWiki(Cognition / Devin) | Free web page + MCP | 2025-04 | Indexed over 30k repos from launch; the open-source clone AsyncFuncAI/deepwiki-open has 18.4k stars | Any public repo can answer questions about its code and docs; a free feature at big-company level |
| HF Cookbook "Simple RAG on GitHub issues" | Tutorials | 2024 | Official HF tutorial | A starter example on the same topic, with no evaluation |
| yanchuk/github-issues-rag | Open source | Unknown | Fewer than 50 stars (not in our repo list) | A personal project with the same name and topic, with no evaluation or judge audit |
| ragas(vibrantlabsai/ragas) | Open-source evaluation frameworks | 2023-05 | 16.3k stars, +58 over 7 days | A library of general RAG evaluation metrics; issue-rag is not a framework, but one concrete experiment |
| DeepEval(confident-ai) | Open-source evaluation frameworks | 2023-08 | 19.5k stars, +146 over 7 days | Same as above, and it comes with LLM-as-judge built in |
| promptfoo | Open source | 2023-04 | 27.1k star | Testing and red-teaming for prompt / RAG / agent |
| Arize Phoenix / TruLens / RAGChecker(Amazon) | Open source | 2020–2024 | 12.3k / 3.6k / 1.1k star | Observability and fine-grained diagnosis; RAGChecker is a companion to a paper |
| vectara/open-rag-eval | Open source | 2024-11 | 423 star | Evaluation that doesn't require reference answers |
| knowflow-ai/KnowEval | Open source | 2025-11 | 53 star | End-to-end RAG evaluation for Chinese enterprise documents, likewise a small "methodology"-type repo |
- What truly sets it apart is the approach, not the features: it uses maintainer replies as the gold answers; it excludes the source issue during evaluation; it classifies "confidently wrong answers" as incorrect; it audits the judge; and it reports only intervals. These methods aren't new in evaluation circles (RAGAS and papers on LLM-as-judge bias cover them), but few personal projects have organized them into a README with complete data that others can verify.
- This differentiation is not why it got stars. Larger projects with more solid methodology are gaining at +58 to +146 per week, while KnowEval, with similar methodology, has only 53 stars. issue-rag has 339 stars in 12 days, with no real human discussion at all, which doesn't match the content quality.
5. Growth tactics broken down
- Buying stars / forks (the main method, strong inference): the boosting network delivers stars evenly across days, about 300 stars a week, plus 1 forks a day so the fork-to-star ratio looks normal (1.8%). The cost is an external service fee; the return is the "12 days, 300+ stars" figure on a résumé, plus screenshots of "it's been covered" from automatic aggregation.
- README written to be picked up (real, and effective): the title, intro, and first screen are organized around "a counterintuitive number + a range + a verifiable results page + an upfront admission of limitations," in both Chinese and English. That's why the AI news sites and @xqliu were willing to relay it from the README with almost no rework (reveneau's original text says "The author's own README is the source for every number below").
- Automatic aggregators and charts picking it up (passive): amazingindex, reveneau, aiqipage, trendshift, and the Threads/LinkedIn template accounts all pick repos by velocity.
- Not found: the author posted nothing on X / Xiaohongshu / Zhihu / V2EX / HN / Reddit (no social account linked to the author; search found nothing either).
6. How many people know about it
Not even small circles discuss it. Everything findable is automatic aggregation plus one low-engagement tweet (584 views, 3 likes). There are no human discussion threads, no issues, no PRs, and the watcher count is 0. It has no revenue involved.
Copycats: no new repo found copying the "issue RAG + grader audit" approach. But the same star-farming network is packaging 20 to 30 small Chinese repos at once, and "star farming + resume projects" is already a playbook many people are running.
7. What to borrow
- The README is worth studying: a counterintuitive number as the hook; ranges reported, not the best-looking figure; a static results page you can check line by line; limitations stated up front. READMEs like this get relayed word for word by AI news sites and KOLs.
- The "auto-pick by growth rate" aggregators (amazingindex, reveneau, trendshift, template-repost accounts) can be triggered by a manufactured growth spike. That's both a distribution tactic and a blind spot in our data recall: looking only at star curves, you can't tell faked from real.
- A suggested rule for recall: check the fork accounts. Accounts with usernames like "word+word+2 random characters", 0 followers, and public repos that are mostly forks get flagged as suspicious. Repos starred by the same set of accounts often have identical curves, so you can pull a whole batch of client repos by "shared stargazer accounts".
8. Still unclear
- The stars themselves can't be verified one by one: GitHub hides the stargazer list. "Bought" is inferred from fork accounts, curve fingerprints shared with the same network's repos, and zero real-human signals.
- We don't know whether the author bought them himself or someone bought them for him (a bootcamp, a resume-packaging service).
- Who "ccb" is, and how much of the repo is AI-generated, can't be confirmed; the NOTES file itself says it's an "AI-assisted project".
- Will GitHub clean up a chunk of this later? Repos on the same network have already been taken down. 11-06/07
Sources
- Repo, commits, fork list, account profiles: github.com/zaxiu88888/issue-rag (GitHub API, checked live 2026-11-13)
- NOTES.zh-CN.md "interview cheat sheet": github.com/zaxiu88888/issue-rag/blob/main/NOTES.zh-CN.md
- Example star-farming accounts: github.com/bronzer-cornel1f, github.com/dragmailer1u, github.com/swath-bastion-8w
- Star-count curves of the same network's repos: our github repo_day / repo_profile (2026-11-01 to 11-13)
- @xqliu tweet: x.com/xqliu/status/2120347028219719706 (checked on X: ❤3 🔁0 💬5 👁584)
- reveneau (AI-generated news, 11-07): reveneau.com/ainews/issue-rag-measures-rag-vite-95-items-aud…
- amazingindex daily (11-07): www.amazingindex.com/article/c9859ef4562708ab178b1e692a7ca96…
- Threads template post: www.threads.com/@arthi__rajendran/post/DfcN4dpFS1O/; LinkedIn: www.linkedin.com/posts/arthirajendran_github-zaxiu88888issue…
- AI flag page: www.aiqipage.com/news/4770; trendshift: trendshift.io/topics/rag
- Competitors: dosu.dev/oss, docs.kapa.ai/knowledge-sources/connectors/github-issues, cognition.com/blog/deepwiki, github.com/kubernetes/website/discussions/57009; star counts from our github data