3.5 KiB
Index
This module contains the components that make up the search index.
It exposes an API for querying the index, and contains the logic for ranking search results. It does not parse the query, that is the responsibility of the search-query module.
Indexes
There are two indexes with accompanying tools for constructing them.
-
index-reverse is code for
word->documentindexes. There are two such indexes, one containing only document-word pairs that are flagged as important, e.g. the word appears in the title or has a high TF-IDF. This allows good results to be discovered quickly without having to sift through ten thousand bad ones first. -
index-forward is the
document->wordindex containing metadata about each word, such as its position. It is used after identifying candidate search results via the reverse index to fetch metadata and rank the results.
Additionally, the index-journal contains code for constructing a journal of the index, which is used to keep the index up to date.
These indices rely heavily on the libraries/btree and libraries/array components.
Result Ranking
The module is also responsible for ranking search results, and contains various heuristics for deciding which search results are important with regard to a query. In broad strokes BM-25 is used, with a number of additional bonuses and penalties to rank the appropriate search results higher.
Central Classes
Domain Ranking
The module contains domain ranking algorithms. The domain ranking algorithms are based on the JGraphT library.
Two principal algorithms are available, the standard PageRank algorithm, and personalized pagerank; each are available for two graphs, the link graph and a similarity graph where each edge corresponds to the similarity between the sets of incident links to two domains, their cosine similarity acting as the weight of the links.
With the standard PageRank algorithm, the similarity graph does not produce anything useful, but something magical happens when you apply Personalized PageRank to this graph. It turns into a very good "vibe"-sensitive ranking algorithm.
It's unclear if this is a well known result, but it's a very interesting one for creating a ranking algorithm that is focused on a particular segment of the web.
Central Classes
- PageRankDomainRanker - Ranks domains using the PageRank or Personalized PageRank algorithm depending on whether a list of influence domains is provided.
Data sources
- LinkGraphSource - fetches the link graph
- InvertedLinkGraphSource - fetches the inverted link graph
- SimilarityGraphSource - fetches the similarity graph from the database
Note that the similarity graph needs to be precomputed and stored in the database for the similarity graph source to be available.