README.md

quora-scraper

Scrapes Quora answers — off a topic feed, a question page or a profile — into JSON or CSV with a stable column schema, a run-metadata sidecar, and exit codes that tell "blocked" from "empty" from "partial".

Works on all twenty-four of Quora's language sites, which are twenty-four different catalogues: www.quora.com plus es., fr., de., it., jp., id., pt., hi., nl., da., fi., no., sv., mr., bn., ta., ar., he., gu., kn., ml., te. and pl.quora.com.

Read this first if you only read one thing. Which URL you point it at decides what you get back, and the difference is not cosmetic. Quora renders none of its content on the server; what it does send is its own GraphQL results, inlined into the page — and only some page kinds carry the answers.

you ask for	rows	upvotes, views, created_at, full text
a question page	its answers	yes — from the inlined payload
a profile	that author's answers	yes, for the ones the payload covers
a topic feed	the topic's answers	no — card only, body truncated to 3 lines

A topic page's answers arrive over a later XHR, so its inlined payload holds 0 answer objects (measured, two topic captures in two languages). The data_source column on every row says which view built it, and the run says so at startup.

---

Install and run

git clone https://github.com/2scraper/quora-scraper && cd quora-scraper
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt -r requirements-playwright.txt
playwright install chromium

python3 playwright_scraper.py \
  --url "https://www.quora.com/What-is-machine-learning-4" \
  --pages 3 --out ml

That writes ml.json, ml.csv and ml.meta.json.

chromium, not chrome — and unlike a sibling repo in this family that is not a compromise. See below.

Install exactly one engine. The three engines' pins are mutually unsatisfiable (playwright and pyppeteer disagree on pyee; pyppeteer and selenium on urllib3). Use a virtualenv per engine if you need more than one.

---

The six things that will surprise you

1. --pages counts scroll batches, not addresses

Quora has no per-page URL in any mode. There is no link[rel=next], no pagination control, no numbered anchors — and ?page=2 does not fail. It is ignored, and the feed comes back with its first items again.

That is the dangerous part. A scraper built on the obvious convention would fetch ?page=2, find no answer it had not already seen, conclude the listing was exhausted and report a complete run holding one batch. So:

• a "page" here is one settled scroll batch;
• the listing ends when a batch adds no sku the run has not already seen;
• --concurrency above 1 is ignored, with a warning that gives the reason. The run continues with one worker. Batch 5 exists only inside the browser that scrolled through batches 1-4; there is nothing to hand a second worker. Run several topics in parallel instead, one process each.

Each batch re-parses the whole feed, so rows_new_per_batch in the sidecar is the number worth reading rather than the raw per-batch count.

2. A topic run has no upvotes, views or creation times

Not a bug and not a threshold — it is what the site sends. On a topic page:

inline GraphQL payloads in the served HTML   3
Answer objects in them                       0

against a question page's 16 payloads carrying 5 answers with their numeric ids, full text, numUpvotes, numViews, numShares and an epoch-microsecond creationTime.

So upvotes, views, shares, comments, created_at, answer_id and question_id are null on every topic row, and text there is what the card rendered — which Quora truncates to three lines. text_chars makes that visible at a glance and data_source says which view built the row.

If you want the counts, point it at a question or a profile.

3. The block is Cloudflare, and the address's recent rate is what decides

What a refusal looks like: HTTP 403, cf-mitigated: challenge, <title>Just a moment...</title>, about 6 KB of body, and cType: 'managed' in the challenge's own config.

Measured 2026-09-15, one DATACENTRE address (a netcup host in Nuremberg), about forty-five fetches over three hours:

window	fetches	served	challenged
first ~20 minutes	14	8	6
everything after	~30	0	all of them

Three things follow:

• A rested address recovers. Early in that window the same URL that answered 403 answered 200 with the full feed on the next attempt about a minute later, from the same address and the same browser. So the first thing to try is --retries 4 --retry-delay 40.
• A busy address does not, and it does not recover quickly. After about thirty fetches every subsequent attempt was challenged, and a 45-minute rest did not clear it — nine further attempts across the next twenty minutes were all refused. Nothing about that address changed except how much it had been fetching.
• The score follows the ADDRESS, not the language site. es.quora.com refused the same address at the same moment www.quora.com did, with three attempts each. Switching hosts is not a way around it.

So --delay is the cheapest lever, --proxy-file is the one that scales, and "it worked an hour ago" is not evidence that it will work now.

And one thing that follows for your wallet, stated precisely because it decides whether a 2Captcha key would help you here. A managed challenge carries no sitekey in the markup — 0 data-sitekey attributes and 0 Turnstile iframes on the one measured. That is a fact about the markup and not about the challenge: Cloudflare passes sitekey, action, cData and chlPageData to turnstile.render() once and keeps nothing, so no static read of the HTML can build a solvable task, however careful it is. Those arguments can be captured with an init script installed on the context before any page script runs, and 2Captcha solves what comes out as TurnstileTaskProxyless.

This repo does not implement that interception. The reason is the measurement above rather than any limit of the solver: what gates this site is the address's recent request RATE, so a solved challenge buys one page from an address that is about to be challenged again, while --delay and --proxy-file change the thing actually being measured. So this scraper never attempts a solve on a challenge and never charges you for one. If you need the interception on a site where a token is worth buying, foodpanda-scraper in this family implements it.

If a run exits 3, the interstitial is saved beside your output as <out>_page<N>_debug.html with a screenshot next to it.

4. How many answers you get is decided at LOAD, not by scrolling

Four loads of the same question URL, minutes apart, same machine, same browser build:

cards at first paint	5	13	5
after scrolling to the bottom and waiting 20s	5	13	259

In the short sessions the scroll worked — scrollY reached the document's end and stayed there — and Quora simply never fetched more. So a run that comes back with six rows out of a question's stated 253 is not a broken scroll, and no amount of extra patience will fix it.

Re-run it. A fresh browser re-rolls the variant. The run warns when it sees this, rather than quietly reporting a short result as a complete one.

5. A Space or a Quora Session answer lives on its own subdomain

Four of one profile's eighteen answer permalinks were on {session}.quora.com, with no /answer/ segment at all — which is the same URL shape as a question page on www. That is why the card anchor here is Quora's own a.answer_timestamp class rather than a URL pattern: no pattern can tell those two apart.

On such a host the permalink is the question's path plus a per-answer suffix, and Quora's own payload reports a third value again:

permaUrl       …quora.com/How-important-is-…-in-Machine-Learning-1
question.url   …quora.com/How-important-is-…-in-Machine-Learning
question.slug  How-important-is-…-in-Machine-Learning-5

So question_url is exact when the payload reached the row and is the permalink itself otherwise — a stated approximation rather than a guess dressed up as a reading.

6. Traps that look like bugs

• answer_count is much larger than the rows you got. One question page renders twelve answers and reports 253. That is Quora rendering a fraction and extending it on scroll, not a gap this run failed to close — which is why it sits beside the row count in the sidecar rather than being subtracted from it.
• author_credential is null on some rows. 28 of 30, 27 of 30, 14 of 18 and 10 of 13 across four captures. The gap is authors who have not written one.
• date_text is a display string, never a date. "Aug 14", "2y", "3 años" — in the page's own language, and "2y" cannot be resolved to a day. created_at is the real timestamp and comes only from the payload.
• language is null on a Space or Session row. Those hosts are not language sites, and their language is whatever their writers use.
• A run that finds nothing writes nothing, so a bad run cannot overwrite last night's good file. --allow-empty is the opt-out: it writes the empty file, and the exit code still says why there were no rows (4 for a genuinely empty feed, 3 for a block, 5 for a run that never got its content).

---

Output

One row per answer. Same columns in JSON and CSV, same order.

{
  "source": "www.quora.com",
  "scraped_at": "2026-09-15T09:14:02.881293+00:00",
  "url": "https://www.quora.com/What-is-machine-learning-4/answer/Someone-1",
  "sku": "www.quora.com/What-is-machine-learning-4/answer/Someone-1",
  "title": "What is machine learning?",
  "question_url": "https://www.quora.com/What-is-machine-learning-4",
  "question_id": "4256391",
  "answer_id": "1477743744691014",
  "answer_count": 253,
  "author": "Someone",
  "author_url": "https://www.quora.com/profile/Someone-1",
  "author_credential": "former Data Scientist at a large company",
  "author_is_verified": false,
  "text": "Hey, future ML engineers!! …",
  "text_chars": 2150,
  "upvotes": 125,
  "views": 1733,
  "shares": 0,
  "comments": 2,
  "created_at": "2024-03-07T08:31:46.062197Z",
  "updated_at": null,
  "date_text": "2y",
  "is_machine_answer": false,
  "is_translated": false,
  "language": "en",
  "category": "What-is-machine-learning-4",
  "mode": "question",
  "data_source": "dom+inline",
  "page": 1,
  "position": 3
}

sample_output.json and sample_output.csv are cut from a real run.

There is no price, currency, rating or in_stock column, unlike the rest of this scraper family. Quora has none of those things, and six columns null on every row of every run of every mode is worse than six missing ones. upvotes is a count rather than a score and is not written under rating.

Exit codes

code	meaning
0	rows written
1	crash
2	bad usage
3	blocked — Cloudflare's challenge survived the retries
4	ran fine, the site served the feed, and it held nothing
5	the content was never obtained: a remote API error, a navigation timeout, a dead proxy, a proxy that refused the credential or the balance before the run began (stop_reason: proxy_auth_failed / proxy_balance_exhausted), the proxy budget spent before any row (proxy_budget_exceeded), a refused batch, an address that answered 404, or a page that held answers and parsed to none (stop_reason: parser_found_nothing, a parser bug rather than an empty feed)
6	partial — see stop_reason in the sidecar (proxy_budget_exceeded when --proxy-budget-mb stopped it)

<out>.meta.json records status, stop_reason, which batches failed by number, the scroll trace, rows_new_per_batch, answers_available and inline_payload_rows. A failed run writes no sidecar, so a "failed" file can never sit beside good data. The one exception is a run the proxy pre-check stopped: it writes <out>.latest_attempt.meta.json, a separate file, and leaves <out>.meta.json alone. A run through a local proxy always writes that file too, with its proxy traffic, rows or not.

Diffing two runs

python3 diff_runs.py --old ml.2026-09-01.json --new ml.2026-09-08.json

Keyed on sku. The bucket to know about here is source_changed: a topic row has null counts and the same answer read off its question page has real ones, so diffing the two would report every counter as having appeared from nowhere. That goes in its own bucket and --fail-on-change ignores it.

--price-tolerance-pct keeps the family's name and here means a COUNT tolerance. Unlike in most of this family it has a real use: Quora's view and upvote counts are live, so a monitor watching for an answer taking off wants a threshold — while one watching for an edit wants text_chars, which the tolerance never absorbs.

---

Engines

script	driver	notes
playwright_scraper.py	Playwright	primary. The only one with --fingerprint and --locale.
selenium_scraper.py	Selenium + chromedriver	Cannot authenticate a proxy, and cannot use an authenticated CDP endpoint.
puppeteer_scraper.py	pyppeteer	Parity engine. pyppeteer is effectively unmaintained and its own README points at Playwright.
scraper_api_client.py	2Captcha Scraper API	One HTTP request, no local browser. The richest path per row on a question URL and useless on a topic one — see below.

All three browser engines share product_parser.py, page_flow.py, output_writer.py, proxy_pool.py and captcha_solver.py, so they agree on rows, exit codes, run status and whether a run spends money. The offline suite binds every shared-module call in every engine against the callee's real signature, because two engines in a sibling repo once called one with a keyword where the parameter was positional and both crashed on their first fetch.

Known limits, stated rather than left to be discovered:

• Selenium cannot use an authenticated remote CDP endpoint. Playwright's connect_over_cdp and pyppeteer's browserWSEndpoint take a full ws://user:pass@host:port and authenticate on the WebSocket upgrade; chromedriver's debuggerAddress takes a bare host:port with nowhere to put a password.
• Selenium cannot authenticate a proxy at all. Credentials are stripped with a warning rather than silently half-working.
• --fingerprint and --locale exist on Playwright and Selenium and not on pyppeteer. The offline suite asserts that difference, so closing it needs a README edit rather than a quiet patch.

Proxy traffic savers

A proxy is billed by traffic, so a run through --proxy / --proxy-file / QUORA_PROXY does two things to avoid paying for bytes it does not use:

• One pre-check before any browser starts. A CONNECT through the proxy and a HEAD /robots.txt on the target host. Chromium reports a proxy's HTTP 407 as a navigation timeout, so a wrong or exhausted credential used to spend every retry before ending as page_load_timeout. Now a 407 (the gateway refused the credential, or the traffic balance is exhausted) or a 402 (proxy traffic/balance exhausted) ends the run in seconds with exit 5 and <out>.latest_attempt.meta.json saying status: failed, stop_reason: proxy_auth_failed (or proxy_balance_exhausted). Any previous output and its .meta.json are left as they were. The message names the variable and never prints the credential. If the proxy does not answer at all, the run goes on as before and the browser makes its own attempt. No pre-check runs without a proxy or with --cdp-endpoint.
• Images, video/audio and web fonts are not downloaded through a local proxy: all three engines abort those requests (Playwright by resource type, pyppeteer through CDP's Fetch domain, Selenium by file extension with CDP Network.setBlockedURLs). Nothing this scraper reads is in them. Measured 2026-10-20 on one question page, with no proxy and blocking forced on: 262 rows either way, the same 30 fields, and the same non-null count in every field. --load-media (or QUORA_LOAD_MEDIA=1) loads them. Without a proxy, or over --cdp-endpoint, nothing is blocked.
• Every run is metered and capped. --proxy-budget-mb (default 10, 0 = no cap, also QUORA_PROXY_BUDGET_MB). Once more than that has gone through the proxy, further requests are refused, the run stops before its next navigation or scroll, keeps the rows it has and exits 6 (or 5 with no rows), stop_reason: proxy_budget_exceeded. It never carries on past the cap. The sidecars (<out>.meta.json, and <out>.latest_attempt.meta.json, which every metered run writes) carry proxy_bytes, proxy_budget_bytes (null with no cap) and proxy_budget_exceeded, and the log ends with Proxy traffic: X.X MB of Y MB budget.

How the bytes are counted, per engine: Playwright sums request.sizes() (request headers + body, response headers + body) of every finished or failed request; pyppeteer reads CDP Network.loadingFinished.encodedDataLength (bytes received, headers included) plus each request's line, headers and body from Network.requestWillBeSent; Selenium reads the same CDP events from chromedriver's performance log after each page and scroll. Only http(s) requests and responses that came over the network are counted: Chrome's own chrome://, data: and blob: loads, cache hits and Chrome's error page count nothing. The figure is an estimate of the payload through the tunnel: TLS, the CONNECT exchange and connections the browser opens for itself are not visible to any of the engines, and add a few percent. Measured 2026-10-21 against a local stub proxy that counted every byte it carried (a 60 kB page, a 30 kB stylesheet, a 90 kB script): Playwright within 0.2%, pyppeteer 0.6%, Selenium 0.7%. A megabyte here is 10^6 bytes. Nothing is metered without a proxy or over --cdp-endpoint.

---

The four paths, measured against each other

All four read the same rows into the same columns. What differs is how many rows and how much of each, and on this site the difference is larger than usual — measured 2026-09-15 on the same question URL:

path	rows	payload-backed	cost
a browser engine, 1 scroll batch	260	6	free
Scraper API, one request	6	6 of 6	$0.0005

The API returns the SERVED response, so every row it produces comes out of Quora's inlined payload with its upvote count, view count, creation time, numeric ids and full text. A browser engine reaches forty times as many rows and reads almost all of them off cards, which carry none of those columns.

On a TOPIC url the API returns nothing at all — 99,833 bytes, 3 inline payloads, 0 answer objects, 0 cards — because a topic's answers arrive over a later XHR. --wait-element 'a.answer_timestamp' does not help: the same URL took sixteen seconds instead of three and came back with the same shell.

So: the API for a question read in depth-of-columns, a browser engine for a topic feed or for breadth of rows.

Do you need any of the paid products?

Not to get started. Playwright's own bundled Chromium, from an ordinary DATACENTRE address, with no key and no proxy, was served HTTP 200 and the full feed on eight of the first fourteen fetches on 2026-09-15, and the six refusals cleared on the next attempt. That is a stronger result than it sounds: a datacentre range is the first thing any bot manager scores, and this one was served anyway.

To keep going, yes. Everything after the first fourteen fetches from that same address was challenged — about thirty attempts over three hours, a 45-minute rest included, and a second language host made no difference. Nothing about the address changed except how much it had been fetching.

So what the paid products buy here is rate: many addresses is how you fetch a lot of Quora, and one address plus a generous --delay is how you fetch a little of it for free.

product	what it buys here
Proxies	spreads the request rate across addresses, which is the real limit
Scraping Browser API	a remote browser over CDP, with persistent profiles — no browser infrastructure of your own
Fingerprints	a consistent device identity across runs
Scraper API	a question's answers in one request, with every column populated, for $0.0005 — measured. Not a topic feed: see the table below.
Captcha solving	nothing in this repo today, because this repo does not implement the Turnstile interception that Quora's refusal needs. The refusal is a managed Cloudflare challenge whose sitekey is not in the markup; its arguments can be captured from turnstile.render() and solved as TurnstileTaskProxyless (see the captcha section above), but the solver here is wired for reCAPTCHA only. At most one solve is bought per page.

All four sit behind one key. Put it in .env (see .env.example) rather than on a command line — a secret in argv is readable by anything that can run ps and lands in shell history.

python3 env_config.py     # prints what was picked up, WITHOUT secrets

---

What Quora publishes, and where this reads it from

No JSON-LD anywhere: 0 application/ld+json blocks on topic, question, profile, answer-permalink and language pages alike. And no server-rendered content: 0 occurrences of Quora's own q-box class in the raw response body of six captures, against 385-1165 in the hydrated document.

What the server does send is its own query results:

window.ansFrontendGlobals.data.inlineQueryResults.results["<hash>"]
    .push("{\"data\":{…}}")

That is this site's structured data, and where it covers a row it is richer than anything rendered. Where it does not, the rendered card is read through Quora's own test and logging hooks — a.answer_timestamp, .puppeteer_test_question_title, .spacing_log_answer_content, .spacing_log_answer_header — which are semantic names rather than build hashes and were stable across every capture and every language host.

Measured coverage on four captures, two languages:

capture	rows	author	body	credential	payload-backed
topic (en)	30	30	30	28	0
topic (es)	30	30	30	27	0
question (en)	13	13	13	10	5
profile (en)	18	18	18	14	3

---

Testing

python3 smoke_test.py        # offline, no network, no engine library needed
pytest                       # the same checks, wrapped as one pytest test

The suite passes with no engine installed at all — every engine import is guarded and the skip is reported, because "skipped, engine absent" reads exactly like a real import error. CI's engine-smoke job installs each engine in its own virtualenv and fails if any group reports a skip.

Fixtures are cut from real captures by make_fixtures.py, which proves each one parses identically to its untrimmed original, column for column. Author names, profile slugs, credentials, answer bodies and cited links are replaced with placeholders first; question titles, ids, counts, timestamps and Quora's own markup stay verbatim.

canary.yml runs two real runs weekly (Mondays) — a question URL for the inline-payload path and a topic URL for the DOM-only one — and asserts that each behaves the way this README says it does. A run that did not parse real rows (blocked, never got its content, partial, or empty) turns the badge red rather than green: from a shared GitHub runner Cloudflare's rate-based challenge does refuse some days, and a QUORA_PROXY secret is what makes those days testable.

Both canary runs pass --proxy-budget-mb 25, above the scraper's own 10 MB default, which is unchanged. Measured 2026-10-21 through QUORA_PROXY: a question run (3 pages) reached the 10 MB cap with images, video/audio and fonts already blocked, so a 10 MB canary would stop on a healthy page. The step summary shows each run's traffic against the budget recorded in its sidecar, and a run that hits 25 MB still fails the job.

---

Scope and licence

This scrapes public pages: topic feeds, question pages and profiles, as an anonymous reader sees them. It does not log in, does not touch anything behind Quora's sign-in wall, and /search is refused outright with that reason — Quora answers an anonymous search request with the language home page rather than with results.

Respect the site's terms and robots policy, and keep your request rate low — the measurements above are as much about being a good citizen as about getting data.

MIT. See LICENSE, CONTRIBUTING.md, SECURITY.md and TROUBLESHOOTING.md.