Scalpel

Scalpel is a web scraping library inspired by libraries like Parsec and Perl's Web::Scraper. Scalpel builds on top of TagSoup to provide a declarative and monadic interface.

There are two general mechanisms provided by this library that are used to build web scrapers: Selectors and Scrapers.

Selectors

Selectors describe a location within an HTML DOM tree. The simplest selector, that can be written is a simple string value. For example, the selector "div" matches every single div node in a DOM. Selectors can be combined using tag combinators. The // operator to define nested relationships within a DOM tree. For example, the selector "div" // "a" matches all anchor tags nested arbitrarily deep within a div tag.

In addition to describing the nested relationships between tags, selectors can also include predicates on the attributes of a tag. The @: operator creates a selector that matches a tag based on the name and various conditions on the tag's attributes. An attribute predicate is just a function that takes an attribute and returns a boolean indicating if the attribute matches a criteria. There are several attribute operators that can be used to generate common predicates. The @= operator creates a predicate that matches the name and value of an attribute exactly. For example, the selector "div" @: ["id" @= "article"] matches div tags where the id attribute is equal to "article".

Scrapers

Scrapers are values that are parameterized over a selector and produce a value from an HTML DOM tree. The Scraper type takes two type parameters. The first is the string like type that is used to store the text values within a DOM tree. Any string like type supported by Text.StringLike is valid. The second type is the type of value that the scraper produces.

There are several scraper primitives that take selectors and extract content from the DOM. Each primitive defined by this library comes in two variants: singular and plural. The singular variants extract the first instance matching the given selector, while the plural variants match every instance.

Example

Complete examples can be found in the examples folder in the scalpel git repository.

The following is an example that demonstrates most of the features provided by this library. Supposed you have the following hypothetical HTML located at "http://example.com/article.html" and you would like to extract a list of all of the comments.

<html>
  <body>
    <div class='comments'>
      <div class='comment container'>
        <span class='comment author'>Sally</span>
        <div class='comment text'>Woo hoo!</div>
      </div>
      <div class='comment container'>
        <span class='comment author'>Bill</span>
        <img class='comment image' src='http://example.com/cat.gif' />
      </div>
      <div class='comment container'>
        <span class='comment author'>Susan</span>
        <div class='comment text'>WTF!?!</div>
      </div>
    </div>
  </body>
</html>

The following snippet defines a function, allComments, that will download the web page, and extract all of the comments into a list:

type Author = String

data Comment
    = TextComment Author String
    | ImageComment Author URL
    deriving (Show, Eq)

allComments :: IO (Maybe [Comment])
allComments = scrapeURL "http://example.com/article.html" comments
   where
       comments :: Scraper String [Comment]
       comments = chroots ("div" @: [hasClass "container"]) comment

       comment :: Scraper String Comment
       comment = textComment <|> imageComment

       textComment :: Scraper String Comment
       textComment = do
           author      <- text $ "span" @: [hasClass "author"]
           commentText <- text $ "div"  @: [hasClass "text"]
           return $ TextComment author commentText

       imageComment :: Scraper String Comment
       imageComment = do
           author   <- text       $ "span" @: [hasClass "author"]
           imageURL <- attr "src" $ "img"  @: [hasClass "image"]
           return $ ImageComment author imageURL

Tips & Tricks

The primitives provided by scalpel are intentionally minimalistic with the assumption being that users will be able to build up complex functionality by combining them with functions that work on existing type classes (Monad, Applicative, Alternative, etc.).

This section gives examples of common tricks for building up more complex behavior from the simple primitives provided by this library.

OverloadedStrings

Selector, TagName and AttributeName are all IsString instances, and thus it is convenient to use scalpel with OverloadedStrings enabled. If not using OverloadedStrings, all tag names must be wrapped with tagSelector.

Matching Wildcards

Scalpel has 3 different wildcard values each corresponding to a distinct use case.

anySelector is used to match all tags:

textOfAllTags = texts anySelector
AnyTag is used when matching all tags with some attribute constraint. For example, to match all tags with the attribute class equal to "button":

textOfTagsWithClassButton = texts $ AnyTag @: [hasClass "button"]
AnyAttribute is used when matching tags with some arbitrary attribute equal to a particular value. For example, to match all tags with some attribute equal to "button":

textOfTagsWithAnAttributeWhoseValueIsButton = texts $ AnyTag @: [AnyAttribute @= "button"]

Complex Predicates

It is possible to run into scenarios where the name and attributes of a tag are not sufficient to isolate interesting tags and properties of child tags need to be considered.

In these cases the guard function of the Alternative type class can be combined with chroot and anySelector to implement predicates of arbitrary complexity.

Building off the above example, consider a use case where we would like find the html contents of a comment that mentions the word "cat".

The strategy will be the following:

Isolate the comment div using chroot.
Then within the context of that div the textual contents can be retrieved with text anySelector. This works because the first tag within the current context is the div tag selected by chroot, and the anySelector selector will match the first tag within the current context.
Then the predicate that "cat" appear in the text of the comment will be enforced using guard. If the predicate fails, scalpel will backtrack and continue the search for divs until one is found that matches the predicate.
Return the desired HTML content of the comment div.

catComment :: Scraper String String
catComment =
    -- 1. First narrow the current context to the div containing the comment's
    --    textual content.
    chroot ("div" @: [hasClass "comment", hasClass "text"]) $ do
        -- 2. anySelector can be used to access the root tag of the current context.
        contents <- text anySelector
        -- 3. Skip comment divs that do not contain "cat".
        guard ("cat" `isInfixOf` contents)
        -- 4. Generate the desired value.
        html anySelector

For the full source of this example, see complex-predicates in the examples directory.

Generalized Repetition

The pluralized versions of the primitive scrapers (texts, attrs, htmls) allow the user to extract content from all of the tags matching a given selector. For more complex scraping tasks it will at times be desirable to be able to extract multiple values from the same tag.

Like the previous example, the trick here is to use a combination of the chroots function and the anySelector selector.

Consider an extension to the original example where image comments may contain some alt text and the desire is to return a tuple of the alt text and the URLs of the images.

The strategy will be the following:

to isolate each img tag using chroots.
Then within the context of each img tag, use the anySelector selector to extract the alt and src attributes from the current tag.
Create and return a tuple of the extracted attributes.

altTextAndImages :: Scraper String [(String, URL)]
altTextAndImages =
    -- 1. First narrow the current context to each img tag.
    chroots "img" $ do
        -- 2. Use anySelector to access all the relevant content from the the currently
        -- selected img tag.
        altText <- attr "alt" anySelector
        srcUrl  <- attr "src" anySelector
        -- 3. Combine the retrieved content into the desired final result.
        return (altText, srcUrl)

For the full source of this example, see generalized-repetition in the examples directory.

Operating with other monads inside the Scraper

ScraperT is a monad transformer scraper: it allows lifting m a operations inside a ScraperT str m a with functions like:

-- Particularizes to 'm a -> ScraperT str m a'
lift :: (MonadTrans t, Monad m) => m a -> t m a

-- Particularizes to things like `IO a -> ScraperT str IO a'
liftIO :: MonadIO m => IO a -> m a

Example: Perform HTTP requests on page images as you scrape:

Isolate images using chroots.
Within that context of an img tag, obtain the src attribute containing the location of the file.
Perform an IO operation to request metadata headers from the source.
Use the data to build and return more complex data

-- Holds original link and data if it could be fetched
data Image = Image String (Maybe Metadata)
  deriving Show

-- Holds mime type and file size
data Metadata = Metadata String Int
  deriving Show

-- Scrape the page for images: get their metadata
scrapeImages :: URL -> ScraperT String IO [Image]
scrapeImages topUrl = do
    chroots "img" $ do
        source <- attr "src" "img"
        guard . not . null $ source
        -- getImageMeta is called via liftIO because ScrapeT transforms over IO
        liftM (Image source) $ liftIO (getImageMeta topUrl source)

For the full source of this example, see downloading data

For more documentation on monad transformers, see the hackage page

Explicit error handling

ScraperT is an instance of MonadError which allows you to throw errors from within parsing code to stop parsing and return an error.

When doing error handling in this way, there are 3 cases to consider:

An explicitly thrown error
A failed scraping without a thrown error
A valid result

This can be implemented for String valued errors as follows:

type Error = String
type ScraperWithError a = ScraperT String (Either Error) a

scrapeStringOrError :: String -> ScraperWithError a -> Either Error a
scrapeStringOrError html scraper
        | Left error    <- result  = Left error
        | Right Nothing <- result  = Left "Unknown error"
        | Right (Just a) <- result = Right a
    where
    result = scrapeStringLikeT html scraper

To add explicit erroring you can use the <|> operator from Alternative to throw an error when something fails:

comment :: ScraperWithError Comment
comment = textComment <|> imageComment <|> throwError "Unknown comment type"

With this approach, when you throw an error it will stop all parsing. So if you have an expression a <|> b and there is a nested throwError in a, then the parsing will fail. Even if b would be successful.

For the full source for this approach, see error-handling in the examples directory.

Another approach that would let you accumulate errors without stopping parsing would be to use MonadWriter and accumulate debugging information in a Monoid like a list:

type Error = String
type ScraperWithError a = ScraperT String (Writer [Error]) a

scrapeStringOrError :: String -> ScraperWithError a -> (Maybe a, [Error])
scrapeStringOrError html scraper = runWriter . scrapeStringLikeT

Then to log an error you can use tell:

comment :: ScraperWithError Comment
comment = textComment <|> imageComment <|> (tell ["Unknown comment type"] >> empty)

You can also retrieve the current HTML being parsed with html anySelector and incorporate that into your log message:

logError :: String -> ScraperWithError a
logError message = do
  currentHtml <- html anySelector
  tell ["Unknown comment type: " ++ html]
  empty

comment :: ScraperWithError Comment
comment = textComment <|> imageComment <|> logError "Unknown comment type: "

For the full source for this approach, see error-handling-with-writer in the examples directory.

scalpel-core

The scalpel package depends on 'http-client' and 'http-client-tls' to provide networking support. For projects with an existing HTTP client these dependencies may be unnecessary.

For these scenarios users can instead depend on scalpel-core which does not provide networking support and has minimal dependencies.

Troubleshooting

My Scraping Target Doesn't Return The Markup I Expected

Some websites return different markup depending on the user agent sent along with the request. In some cases, this even means returning no markup at all in an effort to prevent scraping.

To work around this, you can add your own user agent string.

#!/usr/local/bin/stack
-- stack runghc --resolver lts-6.24 --install-ghc --package scalpel-0.6.0
{-# LANGUAGE NamedFieldPuns #-}
{-# LANGUAGE OverloadedStrings #-}

import Text.HTML.Scalpel
import qualified Network.HTTP.Client as HTTP
import qualified Network.HTTP.Client.TLS as HTTP
import qualified Network.HTTP.Types.Header as HTTP


-- Create a new manager settings based on the default TLS manager that updates
-- the request headers to include a custom user agent.
managerSettings :: HTTP.ManagerSettings
managerSettings = HTTP.tlsManagerSettings {
  HTTP.managerModifyRequest = \req -> do
    req' <- HTTP.managerModifyRequest HTTP.tlsManagerSettings req
    return $ req' {
      HTTP.requestHeaders = (HTTP.hUserAgent, "My Custom UA")
                          : HTTP.requestHeaders req'
    }
}

main = do
    manager <- Just <$> HTTP.newManager managerSettings
    html <- scrapeURLWithConfig (def { manager }) url $ htmls anySelector
    maybe printError printHtml html
  where
    url = "https://www.google.com"
    printError = putStrLn "Failed"
    printHtml = mapM_ putStrLn