DEV Community

Coding Agent Index v2.1: scores fell 5 points, 4 configs tie at 58, and cost per task spans 30x

On November 14 Artificial Analysis shipped version 2.1 of its Coding Agent Index, the leaderboard that scores a coding agent as one unit of harness plus model plus effort setting. The short version: nearly everyone lost points, the lead is a four-way tie at 58, and the cost to run one task at the top of the board spans roughly 30x. When the index reset in October I wrote about what the v2.0 reset measures. The v2.1 change is narrower but sharper, and the cost column is the part worth your time if you pick models for a coding agent. What v2.1 changed The index is an equal-weight average of three benchmarks, 322 tasks total, three attempts per task: Component Tasks What it measures How it scores FrontierCode-AA 2.0 132 changes to real open-source repositories maintainer-merge rubric, 0 or 70-100% Terminal-Bench 4.1 66 multi-step terminal work test suite, pass/fail SWE-Atlas-QnA-AA 124 repository question answering binary pass/fail FrontierCode-AA 2.0 is Artificial Analysis' implementation of Cognition's FrontierCode 2.0, released October 27. An attempt that fails any blocking check scores zero. An attempt that passes them all scores between 70% and 100% depending on non-blocking checks. A trajectory scanner zeroes any attempt that went looking for the upstream solution or its merge commit. Terminal-Bench 4.1 keeps 4.0's 66 tasks and environments and replaces the verifiers. Code that runs during grading now runs as an untrusted user in a separate process, and verifiers that rejected valid solutions were corrected. That is reward-hacking hardening, not new tasks. The third component and the weighting did not change. The largest component got harder, and that moved the whole board. The full methodology page is worth a read. All 23 dual-measured configs lost points on the new FrontierCode Per the v2.1 announcement, index scores came down about 5 points on average, driven by the FrontierCode swap. AlphaSignal's read of the underlying data is more specific: all 23 configurations evaluated on both versions scored lower on the new FrontierCode, and the mean FrontierCode score fell 16 points, from 45% to 29%. That is not the models getting worse. FrontierCode 2.0 pulls from 55% more repositories, carries 43% more blocking criteria per task, and its reference solutions modify 67% more files and 89% more lines in codebases that are 25% larger. The share of thoroughness criteria roughly tripled. Thoroughness criteria catch the patch that fixes the reported case and misses the sibling code path, and Cognition reports thoroughness produced the highest blocker failure rate at every reasoning level they analyzed. One configuration barely moved: Codex with GPT-6.1 Terra at max effort stayed within a point of its previous index score. Everyone else ate the reset. This is the second consecutive version change that lowered scores across the board. That is the design. A suite that refreshes before vendors tune against it keeps some discriminating power. A score you saw quoted last month is a different exam. The lead at 58 is two models in four configurations The top score belongs to four configurations, not one product: Claude Code with Fable 5.2 at high effort, Claude Code with Fable 5.2 at xhigh, Claude Code with Opus 5.5 at max, and Devin Fusion CLI with Opus 5.5 Max paired with GPT-6 Luna High. Two details matter more than the rank. First, Fable 5.2 at max scores 56, below its own high-effort result, because more of its attempts hit task time limits. Effort is not a simple dial where more is better. Second, Gemini 4 Carbon at high effort scores 55 in Antigravity CLI. The index deliberately scores the harness and the model together, which is why I trust it more than raw model benchmarks for agent work. We have seen this before from the other side: Hugging Face measured the same model scoring 33% or 62% depending on the harness. The 30x cost spread is the actionable column v2.1 also changed how cost is computed. Artificial Analysis aligned it ...

Powered by Readybot.io

Nov 15, 2026

DEV Community

Gemini CLI v0.68 fixes rootless Podman sandboxing and path boundary checks

Originally published at https://aicoding-guide.com. Gemini CLI v0.68.0 ships with no summary prose at all — the release page is just a list of pull requests. One feature; everything else is a fix. It looks unremarkable, but three of the fixes are about boundaries, and each one is a case where behaviour the docs already describe was not holding in some environment. Key point What you will learn When --userns=keep-id is passed for rootless Podman, and how to override it The fixes to root checks on Windows and to symlink resolution in GlobTool That the new extension validation command and the hooks' SubagentStop are both absent from the docs --userns=keep-id reaches rootless Podman The fix is fix(sandbox): supply --userns=keep-id for rootless podman (#30118). This behaviour is documented, under "Linux UID/GID handling": Gemini CLI automatically handles user permissions in container sandboxes on Linux. When the sandbox runs as your host UID/GID (the default on supported Linux distributions) and you use rootless Podman, Gemini CLI also passes --userns=keep-id. This keeps files that the sandbox creates in mounted directories owned by your host user. To use a different user namespace mode, add your own --userns option to SANDBOX_FLAGS. The promise is that files the sandbox writes into mounted directories stay owned by your host user. If ownership was turning into root under rootless Podman, that promise was not being kept until this release. There are three ways to select the backend, in a fixed order of precedence. Method How Command flag -s or --sandbox Environment variable `GEMINI_SANDBOX=true\ Settings file {% raw %}"sandbox": true inside the tools object of settings.json # Turn on sandboxing with rootless Podman
export GEMINI_SANDBOX=podman
gemini

Pick a different user namespace mode yourself

export SANDBOX_FLAGS="--userns=host"
SANDBOX_SET_UID_GID switches the mapping itself: true forces the host UID/GID, false disables UID/GID mapping. Two fixes to path boundary checks The other two concern which paths count as inside your working area. fix(core): make isWithinRoot case-insensitive on Windows (#29999) — root checks on Windows no longer depend on case, which matters if C:\Project and c:\project were being treated as different places fix(core): resolve symlinked paths canonically in GlobTool (#30042) — GlobTool now resolves symlinked paths canonically Neither behaviour is stated in the documentation The workspace concept itself is documented. The sandbox page says "the sandbox only has access to the current project workspace" and points you at SANDBOX_MOUNTS when something outside it must be reachable. But whether path comparison is case-sensitive on Windows, and how the file tools resolve symlinks, could not be confirmed in the official documentation. The PR titles tell you what changed; no specification-level statement exists. If you need the outside of the workspace genuinely blocked, turn the sandbox on and pass what you need explicitly with SANDBOX_MOUNTS. There is also a fix about git's exclude files: fix(core): honor .git/info/exclude and core.excludesFile in shadow repo (#30034). Here too, no documented statement that .git/info/exclude or core.excludesFile are honored could be found. These are the exclude settings that are documented: Setting What it does context.fileFiltering.respectGitIgnore Respect .gitignore (default true) context.fileFiltering.respectGeminiIgnore Respect .geminiignore (default true) context.fileFiltering.customIgnoreFilePaths Additional ignore file paths, which take precedence over .geminiignore and .gitignore If you want an extra ignore file to count on, naming it in customIgnoreFilePaths is the documented route. The extension validation command and the hooks' SubagentStop The release's only feature is feat(cli): add extension validation command and evaluation tests (#29980). But as of 2026-11-16 no validate subcommand appears in the official documentation. The documented gemini extensions ...

Powered by Readybot.io

Nov 15, 2026

DEV Community

Google News RSS links aren't publisher URLs. Here's how to get the real ones in 2026

TL;DR: Every link in a Google News RSS feed points to news.google.com/rss/articles/CBMi..., not to the publisher. Until July 2024 you could base64-decode the ID and read the URL inside. That stopped working. Today the ID holds an opaque token, and resolving it takes a page fetch plus a call to Google's internal batchexecute endpoint. Loop that over a few dozen articles and people report HTTP 429. You can do it yourself with a free library, or pay about $0.001 per article and get the URL in the same row as the headline. Code for both is below. What the feed gives you Search Google News over RSS: https://news.google.com/rss/search?q=amadeus+api&hl=en-US&gl=US&ceid=US:en
Each has a title, a source, a date and a link like this one: https://news.google.com/rss/articles/CBMihwFBVV95cUxNaUJGcmZlVjJPc18tQkRFd3BUeTBMZGZSOFhQaUZOMGpWR25JQ3M4WnVtR2EwdlBmZWdoTXVQTjRWc3BxcU93NC1XZTl5VTJRYkhFZzFsQW5QeWZoN1JVczljWVIzMU9DdlVWYXBrRUY3c2ZiU003VjUxQm1ManJfTnpSZTlzZkE?oc=5
That's a PhocusWire article about Amadeus. Nothing in the link says PhocusWire. If you want to dedupe by URL, fetch the article text, or group coverage by domain, you need the real URL first. The base64 trick, and why it stopped working For years the answer on Stack Overflow and in blog posts was: take the part after /articles/, base64-decode it, strip a few bytes, and there's your URL. It worked because the ID was a small protobuf with the URL stored in plain text. In late July 2024 Google changed the format. A Stack Overflow question from 25 July 2024 (about 3,800 views) says the base64 method "worked perfectly until Google recently changed the structure". The same day, users of the most popular decoder library opened an issue saying it had stopped working. You can see the change offline. Decode the ID above: import base64

gid = "CBMihwFBVV95cUxNaUJGcmZlVjJPc18tQkRFd3BUeTBMZGZSOFhQaUZOMGpWR25JQ3M4WnVtR2EwdlBmZWdoTXVQTjRWc3BxcU93NC1XZTl5VTJRYkhFZzFsQW5QeWZoN1JVczljWVIzMU9DdlVWYXBrRUY3c2ZiU003VjUxQm1ManJfTnpSZTlzZkE"
raw = base64.urlsafe_b64decode(gid + "=" * (-len(gid) % 4))
print(raw[:4], raw[4:40])
b'\x08\x13"\x87' b'\x01AU_yqLMiBFrfeV2Os_-BDEwpTy0LdfR8XPi'
The protobuf wrapper is still there. Inside it there's now a token starting with AU_yqL, not https://. No amount of decoding turns that into a URL. Only Google can. What decoding takes now The method every working decoder uses today, including the one in my Actor, is the one the Google News web app uses itself: Step Request What you get 1 GET news.google.com/rss/articles/ An HTML page. No redirect header. It carries two attributes, data-n-a-ts (a timestamp) and data-n-a-sg (a signature) 2 POST news.google.com/_/DotsSplashUi/data/batchexecute with garturlreq, the id, the timestamp and the signature A JSON-ish response with the publisher URL So one headline from the feed costs you two more requests to Google. Step 2 can carry several articles in one call. Step 1 can't; it's one page per article. Volume is where it gets hard. In the decoder library's issue #3 (20 comments, August 2024), a user reports a 429 after looping about 50 times, and others confirm "Google News is being strict now". The common advice there is to slow down. Respect that. It's Google's server, and a 429 is a clear signal. Option 1: do it yourself, free googlenewsdecoder implements the two steps above. It's maintained (version 0.2.2, released 18 October 2026) and widely used: 387,792 downloads in the last month on pypistats as of 15 November 2026. pip install googlenewsdecoder
It fits well if you resolve a handful of links a day, from one machine, and can wait between requests. You own the retries and the pacing, and you'll need to update when Google changes the page again. Option 2: a hosted scraper Google News Scraper on Apify reads the RSS feed and resolves publisher URLs in the same run. It batches step 2 (one batchexecute call per 10 articles), retries failures, and runs on Apify's proxies, not your IP. Disclosure: I built this....

Powered by Readybot.io

func ValidateQuery(rawSQL string) error {
tree, err := pg_query.Parse(rawSQL)
if err != nil {
return err // Block malformed SQL by default
}

for _, stmt := range tree.GetStmts() {
    node := stmt.GetStmt()
    if node.GetDropStmt() != nil || node.GetTruncateStmt() != nil {
        return ErrDestructiveQueryBlocked
    }
    if deleteStmt := node.GetDeleteStmt(); deleteStmt != nil {
        if deleteStmt.WhereClause == nil {
            return ErrUnconstrainedDeleteBlocked
        }
    }
}
return nil

Configure a Durable Queue or Fast Streams for Rate-Limiting HTTP-Based Logistics Webhooks To configure a queue for rate-limiting HTTP-based webhook processing in a logistics renewal workflow, choose a durable queue when missing a business deadline is worse than adding a few milliseconds of handoff time. Choose a fast stream only when the downstream contract can tolerate replay, retention, and consumer-management work. The queue is not the rate limiter. It is the shock absorber; the worker-side limiter is the clock that protects the webhook destination. Here is the decision matrix I use before writing a producer: Constraint Durable queue Fast stream Reminder must survive a worker restart Strong fit Possible, with explicit retention and offsets Burst arrival versus fixed webhook quota Pair with a token bucket Pair with a token bucket One consumer should own each job Natural model Requires consumer-group rules Operators need a dead-letter path Built-in pattern in many systems Application convention Lowest possible handoff latency matters Usually not the first choice Better candidate My default is the durable option for renewal reminders. The decision changes when the deadline is soft, the destination is idempotent, and the team already operates stream offsets confidently. That is the whole fork. How should I configure a queue for rate-limiting HTTP-based webhook processing? A logistics renewal reminder has at least three times attached to it: when the carrier account became eligible, when the business wants a reminder sent, and when the webhook receiver will reject a late request. Confusing those timestamps creates a queue that looks healthy while delivering useless work. Store the intended delivery time with the job. Do not use queue arrival as a substitute. A cron task can poll for eligible records every minute, publish them, and exit. That task should not perform the HTTP call itself. A short cron run that discovers 10,000 renewals can otherwise create a burst against a receiver that permits 20 requests per second. The worker needs two independent gates. First, it waits until the job is eligible. Second, it acquires a rate-limit token for the receiver and route. The first gate protects business timing; the second protects an external system. A retry belongs behind both gates, or a failed request can immediately consume the whole retry budget in a tight loop. This is the small contract I want between the scheduler and the worker: type RenewalReminder = {
id: string;
accountId: string;
destination: string;
dueAt: string;
attempt: number;
idempotencyKey: string;
};

function isEligible(job: RenewalReminder, now: Date): boolean {
return Date.parse(job.dueAt) <= now.getTime();
}
The payload is deliberately boring. An immutable job id, an explicit due time, an attempt count, and an idempotency key give the worker enough information to make a decision after a restart. A URL alone does not. The trade-off is intentional: more fields make the contract larger, but they make delayed processing inspectable. No shortcuts. Where should rate limiting live? Put the limit at the narrowest boundary that owns the quota. If a receiver allows 20 requests per second per account, a process-wide limiter is wrong as soon as two accounts share a worker. If the quota is per destination, key the limiter by destination. If the receiver documents a global quota, use a shared limiter or partition work so each worker sees the same budget. A token bucket is a useful mental model: tokens arrive at a fixed rate up to a capacity, and each HTTP attempt spends one. Capacity controls the burst. Refill rate controls the sustained pace. Do not let the queue's concurrency setting pretend to be a rate limit. Ten concurrent requests can still violate a per-second quota, especially when response times vary. The limitation is operational: a shared limiter needs shared state, while a local limiter is easier to run but cannot enforce a global quota. The wo...

Powered by Readybot.io

Nov 15, 2026

DEV Community

I built a decision graph on HydraDB that learns from corrections. Here's what broke.

Agents keep acting on engineering decisions that were reversed years ago. A 48-hour HydraDB build, the bugs I hit, and an honest baseline. Every engineering team writes its decisions down. Then it changes its mind and writes another one. Two years later, search finds the old decision as easily as the new one, and an AI coding agent will happily build on a choice the team reversed long ago. For the HydraDB hackathon I built DecisionGraph to fix that. It links each architecture decision record (ADR) to the records that replaced it. When an agent asks "what did we decide about X?", it walks those links to the decision that still stands and hands back that decision and its history, in about 130 tokens instead of 590. When someone corrects an answer, the correction is written into the graph, so the next question on that topic lands right even in different words. Live app: https://decisiongraph-latest.onrender.com Code: https://github.com/0xNexuz/decisiongraph Video walkthrough: https://youtu.be/ChWL-frv5v8 The problem in one picture Asked "Does each business unit get its own Cloud Platform cluster now?", vector search returns the closest text: a 2019 decision to run one shared cluster, which has been replaced three times since. DecisionGraph starts at the same record, then follows the "replaced by" links to the decision that stands today. The data is about 130 real, public ADRs from three UK government engineering teams: MoJ Cloud Platform, MoJ Modernisation Platform and GOV.UK on AWS (all MIT licensed). The test set is 50 questions I wrote, scored before each correction is applied. How it works Find the decision. Vector similarity, distinctive terms and hints from past corrections each vote for where the question starts. Walk to what replaced it. One HydraDB path query follows trusted "replaced by" links until it reaches a decision nothing has replaced. Learn from the correction. The links on the right path are confirmed, the wrong ones rejected, and the question is stored as a node so the next one like it starts in the right place. Where HydraDB earned its place Trust lives in the path cost Every REPLACED_BY link has a cost: 1 if trusted, 100 if not. Walking only trusted links is then a single algo.MSpaths call with maxCost set to the path length, so one untrusted hop rules the path out: CALL algo.MSpaths({
sourceLabel: 'Decision', sourceProperty: 'key',
sourceValues: ['cloud-platform:12'], targetValues: [...],
relTypes: ['REPLACED_BY'], relDirection: 'outgoing',
maxLen: 5, pathCount: 1, pairwise: false,
weightProp: 'cost', costProp: 'cost', maxCost: 5
}) YIELD path RETURN path
"Follow only what we trust" stopped being a loop in my Python and became one query inside the database. Learning is a write, not a retrain Links come from rules: an ADR's status line naming its replacement, a body that cites an older ADR next to "replace" or "migrate", and so on. Each rule has a trust score (a Beta count). When a correction proves a rule right or wrong, the new trust flows to every link that rule proposed that nobody has judged yet, in one statement: MATCH (x:Decision)-[r:REPLACED_BY]->(y)
WHERE r.rule = $rule AND r.state = 'candidate'
SET r.w = $w, r.cost = $c
The graph that answers the next question is already the corrected one. Memory is part of the graph Every corrected question becomes a Question node with a RESOLVED_TO edge to the right decision, an ABOUT edge to the topic and IN_QUESTION edges from its distinctive terms. "Has someone asked this before?" is a two-hop walk from the new question's terms, not a separate cache. Two cells, two worlds The public demo runs on one HydraDB cell. The "replay all 50 corrections" demo runs on a second cell, so a visitor pressing replay never resets anyone else's graph. Self-hosted for $0 HydraDB is built from source into one Docker image with the app. GitHub Actions builds the image and pushes it to GitHub's container registry, and it runs on Render's free tier. What broke, and h...

Powered by Readybot.io

Nov 15, 2026

DEV Community

Object destructing , Array destructing

Object destructing It unpacks the properties from an object and stores the value into new variable in one step eg: const name = props.name function child({name}) -- collecting them as a separate variable so we don't want to call it as repeatly like props.name, props.age props props... with the help of destructing we reduce it. Array destructing Similar to Object destructing but the only difference is array are taken by the position. eg: const skills = [ "JS" , " React" , "CSS"] const [skills1, skills2, skills3] = skills console.log(skills1) //js

Powered by Readybot.io

Nov 15, 2026

DEV Community

The model was the easy part

A retail brand doesn’t ask how many people walked into a store. It asks why a store with high footfall isn’t selling, which stores deserve more stock before the festive season, and where to open the next one. Answering those questions from a camera feed takes far more than a model that can detect people. I had already shipped multiple features with AI-assisted workflows, and in earlier projects the work usually started and ended with a technical problem. Camera Analytics for retail was different. I was building an ML pipeline for computer vision, and getting the model to work turned out to be the smaller part of the job. The real product was everything around it: understanding what retailers needed, making it work in stores that were all set up differently, and keeping it reliable as features and data grew. Retailers didn’t want a footfall counter Retail brands wanted more than a count of how many people walked in. They wanted to understand customer behaviour: how long people browsed, whether they were greeted, and whether they came in alone or as part of a group. Zones turn a camera view into behaviour: greeting at the entrance, browsing on the floor, and conversion at the billing counter. AI-generated image. Those details fed much bigger decisions. Brands could compare footfall with sales, understand festive shopping trends, and use store performance to plan inventory and new store openings. I spoke with retail clients in India and Poland to understand how they would use this. What surprised me was how far these granular details travelled, well beyond the store floor into decisions across the business. The needs also varied from one brand and store to the next, so understanding the use cases early shaped a product that could support more than one setup. That early focus on real use cases shows in the results: the product converts more than 90% of proofs of concept (POCs) into customers. Every store is a different problem No two stores were set up the same way, and each one brought its own problems. Even a single camera angle changing could affect the setup and the results. If every adjustment needed an engineer, onboarding more stores would become the bottleneck. So I built admin-level controls and an Electron app that handle much of the onboarding work. Someone who understands the system and the model can use it to reconfigure a store remotely, including when a camera angle changes. Documenting the use cases and working through these requirements before building made the biggest difference. The work stopped being about implementing a feature and became about designing how the product would actually be used and maintained. The infrastructure is part of the product Every new pipeline stage or downstream task added load to the system. Before releasing a feature, I needed to know how it affected worker throughput, GPU load and power usage. The goal was to process as much data as possible without throttling. With large volumes of data, each new feature meant benchmarking, training the pipeline, reviewing results and stress-testing the server under high load. Grafana alerts helped me monitor it, but I still wanted to see how the system behaved under real test conditions before trusting it. The next step is moving parts of the pipeline from Python to C++ to reduce the impact of the GIL on parallel execution. Early tests show it is already reducing that bottleneck, though completing the migration will take time. Where the textbook got it wrong Most optimisation advice for GPU inference pipelines is sound in general. On our hardware and under our load, a couple of the obvious moves didn’t do what the textbook promised. Each time, the lesson was the same: measure what is actually happening before deciding what to optimise. For context, the detection model runs on the GPU, while most of the preprocessing happens on the CPU. That split between a fast GPU and a busy host shapes everything below. Detection and tracking on a busy stor...

Powered by Readybot.io

Nov 15, 2026

DEV Community

LLM Weights are Hardcoded Intelligence

Hey everyone 👋, I need to say something upfront that I know some people are going to push back on, and I want to get the disclaimer out of the way before the argument starts. This post is a critique. Not a hate piece, not a dismissal, and not the kind of hot take that evaporates the moment someone asks a follow-up question. I have been building toward this argument across nearly every AI-related post I have written, from LLMs are Useful. LMMs will Break Reality, where I tried to be fair about what LLMs genuinely accomplish, to Training Is an Evil Concept. LMMs Eliminates it Altogether., where I argued that the entire training paradigm carries moral weight that the industry refuses to examine, to Intelligence at Rest, where I unpacked what it actually means for a system to store something worth calling intelligence. Each of those posts was pointing at the same underlying problem from a slightly different angle, and in this post I want to point directly at the center of that problem and name it plainly. The problem is this: LLM weights are hardcoded intelligence. And hardcoded intelligence, no matter how impressively hardcoded it is, is not the same thing as intelligence. It is a fossil of intelligence, a snapshot, a frozen record of what the world looked like at training time, compressed into billions of floating-point numbers that will never change again on their own. I am not saying that is worthless. I am saying that is a fundamentally different thing from what people are claiming it is, and the gap between the claim and the reality is exactly where the most dangerous illusions live. I want to be honest about my own position before anything else, because I think honesty is the only thing that makes a critique worth reading. I am not an established academic with ten published papers on neural network theory. I am a software engineer who has spent years watching the AI industry make increasingly extravagant promises, using language that is increasingly disconnected from the actual mechanics of what these systems do, and I have watched people bet careers and livelihoods and policy decisions on those promises. I wrote in Technology Has Destroyed My Livelihood about how the technology industry dismantled things that mattered to people while calling it progress, and the AI industry is doing something very similar right now with the concept of intelligence itself. It is taking a specific, limited, technically impressive capability, compressing it into a word, and selling the word to the world at a markup that the word cannot support. I am writing this post because I think the gap between the word and the reality deserves to be named clearly, and because the alternative I believe in, which is what we have been building with the LMM project, is real enough and grounded enough to present as an honest alternative. That is what I am going to do in the sections below, and I am going to try to do it without being unfair to the technology I am critiquing. The Weight Is All There Is Let me start at the most basic technical level, because the argument I am making is ultimately a technical one, and I do not want to dress it in philosophy when the engineering is enough. A large language model, in its final deployed form, is a file. It is a collection of tensors, arrays of floating-point numbers, organized according to a specific architecture. For something like GPT-3, that file contains approximately 175 billion individual float values. For GPT-6, the number is estimated to be in the trillions. Each one of those numbers is the result of an optimization process called gradient descent, which ran over hundreds of billions of tokens of text, adjusting each number very slightly with each pass, until the collective behavior of all the numbers together minimized a loss function called cross-entropy, which measures how well the model predicts the next token in a sequence. When that process ends, the numbers are saved. They are written to disk. T...

Powered by Readybot.io

Nov 15, 2026

DEV Community

Ukrainian scans, invoices, Word and Excel to Markdown - 5 Apify Actors on 7 files, only one got every table right

Search the Apify Store for "PDF to Markdown" and you get dozens of Actors with almost the same promise: text, tables, OCR, ready for RAG. None of them tells you what happens with a Ukrainian scan, an Excel file or a scanned invoice whose table has no lines. So I ran five of them on the same seven files and compared every output with the text the files were made from. All results below are from real runs on 15 November 2026. Disclosure: one of the five is mine, Document to Markdown for AI. Every number below comes from a real run, the test files are public and the raw outputs are kept, so you can repeat and check all of it. Short answer Files right (of 7) Invoice without lines: rows exact Reads Word and Excel Price per scanned page ai_admin/document-to-markdown 7 5 of 5 yes $0.004 nidamindllc/doc-ocr 2 4 of 5 no (error) $0.004 lergassy/pdf-text-extractor 1 0 yes $0.0111 fabrikit/document-to-markdown 2 0 yes $0.0105 gochujang/pdf-text-extractor 1 0 (empty) no (PDF only) $0.006 "Right" means the text, the heading and the table with the correct rows and with every number exactly as in the document. The seven files All Ukrainian, all generated from known text, so every character can be checked: File What it is text-uk.pdf PDF with a text layer: heading, two sentences, a 3 x 3 table with prices like 96,50 scan-uk.pdf the same page as a clean scan scan-noisy-uk.pdf the same scan with noise and a slight tilt doc-uk.docx Word document with the text, a list and the table sheet-uk.xlsx Excel workbook: the table with a Total row 1,234.50, and a sheet with the sentences scan-borderless-uk.pdf a scanned invoice: 5 x 6 table without any lines, amounts like 40 530,00 and 2 500 scan-noisy-borderless-uk.pdf the same invoice, noisy and tilted The last two are the hard case and the most common one: most invoices, waybills and bank statements are printed without table lines. The files are public: https://api.apify.com/v2/key-value-stores/FloomuI2DrQCJazY8/records/. The Actors are the most used ones for this search plus one that lists Office formats. Settings: Ukrainian OCR wherever the Actor has a language option (ukr); nidamindllc has no language option. Results, file by file File mine nidamindllc lergassy fabrikit gochujang Text PDF right one OCR typo, table right (HTML) table flattened into the text table as plain lines right Clean scan right right (HTML table) table as broken lines table lost empty Noisy scan right right (HTML table) table as broken lines rows without header empty DOCX right error right right not supported XLSX right, 1,234.50 kept error 1,234.50 became 1234.5 right not supported Invoice, no lines 5/5 rows, 30/30 cells 4/5 rows, 29/30 cells no table, 40 530,00 lost its space no table, all in one line, 96,50 read as 96,0 empty Invoice, no lines, noisy 5/5 rows, 30/30 cells 4/5 rows, 29/30 cells no table, 96,50 read as 96,60 no table, all in one line empty This is what my Actor returned for the noisy invoice without lines - every cell, every number and the № sign as in the document: Рахунок-фактура № 47

Постачальник: ТОВ «Будмайстер», м. Львів. Покупець: ФОП Коваленко О. В. Оплата протягом п'яти банківських днів з дати рахунку.

№ Товар Од. Кількість Ціна Сума 1 Цемент М500 мішок 420 96,50 40 530,00 2 Арматура 12 мм шт. 180 412,00 74 160,00 3 Пісок річковий т 12 650,00 7 800,00 4 Цегла рядова шт. 2 500 9,80 24 500,00

Разом: 146 990,00 грн
The same table also comes as rows of cells in the tables field, ready for a spreadsheet or a database, and the record says how many pages went through OCR, so a pipeline knows which text came from a scan. For comparison, the same line of the invoice from two other Actors: lergassy: 1 | Цемент М500 мішок 420 96,60 40530,00
fabrikit: ... Мо Товар Од. Кількість Ціна Сума 1 Цемент М500 мішок 420 96,50 40 530,00 2 Арматура12мм шт. ...
What stood out gochujang returned empt...

Powered by Readybot.io

**DEV Community** ## Ukrainian scans, invoices, Word and Excel to …

Nov 15, 2026

Pixel

Pixel

Announcements

Coding Agent Index v2.1: scores fell 5 points, 4 configs tie at 58, and cost per task spans 30x

Gemini CLI v0.68 fixes rootless Podman sandboxing and path boundary checks

Pick a different user namespace mode yourself

Google News RSS links aren't publisher URLs. Here's how to get the real ones in 2026

I still can't get an agent to build a good memory. So I build them by hand.

Leaving Lovable Cloud without losing your users

BAS/1C: Platform vs Configuration vs Information Base — A Diagnostic Protocol Before Automation

>-

I Wanted a Simpler Way to Explore Public Snapchat Stories, So I Started Building One

Windows Sentinel 0.5.9 — Copilot desktop tool calls

Compiling Postgres C-Parser to WebAssembly in Go: Building a Layer-4 Guardrail for AI Agents

Stop Treating Scrum Like an Endless Loop of Meetings: Master the 3 Core Artifacts

Godot 4: evenly spaced points along a path (77 stitch holes)

What a 24/7 NOC Actually Monitors — and What It Doesn't

Building a touchscreen macro deck on an ESP32-P4 with LVGL 9.6

Never let the model retype a number: how we read lab reports with two AI passes

Escanear la aplicación publicada: Nuclei y Wapiti en Campus Tareas

Webhook Retry Queue: Delayed Jobs and Idempotency in Node.js

Compare Cloud Storage Providers for Direct Browser Uploads — Node.js Retention Rules

Qué significan las alertas de Nuclei y Wapiti en una app de tareas

Configure a Durable Queue or Fast Streams for Rate-Limiting HTTP-Based Logistics Webhooks

I built a decision graph on HydraDB that learns from corrections. Here's what broke.

Object destructing , Array destructing

The model was the easy part

LLM Weights are Hardcoded Intelligence

Ukrainian scans, invoices, Word and Excel to Markdown - 5 Apify Actors on 7 files, only one got every table right

Announcements

Publishing Permissions

You might also be interested in

Your Data, Your Control

Pixel Updates & Events

Server Updates & Events

Announcements

Coding Agent Index v2.1: scores fell 5 points, 4 configs tie at 58, and cost per task spans 30x

Gemini CLI v0.68 fixes rootless Podman sandboxing and path boundary checks

Pick a different user namespace mode yourself

Google News RSS links aren't publisher URLs. Here's how to get the real ones in 2026

I still can't get an agent to build a good memory. So I build them by hand.

Leaving Lovable Cloud without losing your users

BAS/1C: Platform vs Configuration vs Information Base — A Diagnostic Protocol Before Automation

>-

I Wanted a Simpler Way to Explore Public Snapchat Stories, So I Started Building One

Windows Sentinel 0.5.9 — Copilot desktop tool calls

Compiling Postgres C-Parser to WebAssembly in Go: Building a Layer-4 Guardrail for AI Agents

Stop Treating Scrum Like an Endless Loop of Meetings: Master the 3 Core Artifacts

Godot 4: evenly spaced points along a path (77 stitch holes)

What a 24/7 NOC Actually Monitors — and What It Doesn't

Building a touchscreen macro deck on an ESP32-P4 with LVGL 9.6

Never let the model retype a number: how we read lab reports with two AI passes

Escanear la aplicación publicada: Nuclei y Wapiti en Campus Tareas

Webhook Retry Queue: Delayed Jobs and Idempotency in Node.js

Compare Cloud Storage Providers for Direct Browser Uploads — Node.js Retention Rules

Qué significan las alertas de Nuclei y Wapiti en una app de tareas

Configure a Durable Queue or Fast Streams for Rate-Limiting HTTP-Based Logistics Webhooks

I built a decision graph on HydraDB that learns from corrections. Here's what broke.

Object destructing , Array destructing

The model was the easy part

LLM Weights are Hardcoded Intelligence

Ukrainian scans, invoices, Word and Excel to Markdown - 5 Apify Actors on 7 files, only one got every table right

Announcements

Publishing Permissions

You might also be interested in

مركز المساعدة

Docs

Discord

Telegram DM 👑

Premium Feature

Unlock Premium Support

What's New

UPCOMING FEATURE

Recent Updates

Join the Hype Engine Waitlist

Upcoming: Hype Engine V2

Upcoming: Role Analytics

Recent Updates

Join Spark Tickets Waitlist

Join Spark with Tickets Waitlist

Join the Community Monetization Waitlist

Join X Integration Waitlist

Discord SEO v2

Are you currently using our bot?

Join Role Analytics Waitlist