User Guide

Intended Audience

This guide is intended for developers, data scientists, program managers, or anyone who wants to use Swirl, including searching and customizing SearchProviders.

For background information on Swirl, please review the Swirl Overview.

Terminology

Running a Search

• Open this URL with a browser: http://localhost:8000/ (or http://localhost:8000/galaxy)

If the search page appears, click Log Out at the top, right. The Swirl login page will appear:

• Enter username admin and password password, then click Login.

• Enter a search in the search box and press the Search button. Ranked results appear in just a few seconds:

The Docker version of Swirl does not retain any data or configuration when shut down!

Swirl includes five (5) Google Programmable Search Engines (PSEs) to get you up and running right away. The credentials for these are shared with the Swirl Community.

Using Swirl with Microsoft 365 requires installation and approval by an authorized company Administrator. For more information, please review the M365 Guide or contact us.

Filtering Results by Source

Swirl returns the best results from all available sources by default. To filter results by one or more sources, check one or more of the Source boxes as shown above. Results are instantly filtered to just those sources.

Click Clear All to return to viewing all results.

Sorting Results

Swirl presents results sorted by relevancy by default. To sort by Date or to see the Top Picks from each silo, click on the VIEW BY dropdown and select your preference.

Notes

• Swirl hides results that don’t have a date_published value when sorting by Date.
• As of version 2.5, the DateFindingResultProcessor was added to the Google PSE SearchProvider JSON. It finds a date in a large percentage of results that otherwise wouldn’t have one, and copies the date to the date_published field.

Paging Through Results

To page through results, use the numbered page selectors, or the Prev and Next buttons, above the results list, as shown above.

Starting a New Search

Click the Swirl logo (top left of the page) at any time to reset the Galaxy search form.

Search Syntax

The following table summarizes the current Swirl search syntax options:

AND, OR

AND and OR are passed down to all SearchProviders. Swirl does not verify that results from any SearchProvider comply.

NOT

• NOT is left in queries for SearchProviders with NOT=True in their query_mappings. The NOT applies to all terms following it.

• NOT is rewritten to -term for SearchProviders with NOT_CHAR=- and NOT=False (or not specified). The -term applies to all terms following it.

• NOT and NOT-ed terms are removed from the query for providers that do not have NOT=True in query_mappings.

Swirl scans responses for compliance with NOT statements. Responses that contain NOT-ed content are down-weighted.

Plus/Minus (+/-) Syntax

• A + (PLUS) prefixed to any search term is left in query term and passed down to all SearchProviders.

• A - (MINUS) prefixed to any search term is left in queries for SearchProviders with NOT_CHAR=- configured in their query_mappings. MINUS is rewritten to follow NOT for any source with NOT=True and without NOT_CHAR=- in query_mappings.

• All - (MINUS) prefixed terms are removed from the query for providers that do not have NOT_CHAR=- in query_mappings.

Using Tags to Target SearchProviders

SearchProviders can be given arbitrary Tags that define some scope - topic, type of entity, or whatever concept(s) the source is expected to know about. For example company, or person, or financial. These Tags may be used in searches to direct specific parts of the query to specific sources.

For example: the funding data set included with Swirl has SearchProviders for SQLite3, PostgreSQL and Google BigQuery, all of which contains Tags:

{
    "name": "Company Funding Records (cloud/BigQuery)",
    "connector": "BigQuery",
    ...
    "tags": [
        "Company",
        "BigQuery"
    ]
}

The following query targets the company Tag in these SearchProviders:

electric vehicle company:tesla

For SearchProviders with that Tag, Swirl rewrites the query to just the terms following it. In this case, BigQuery SearchProvider will receive the query:

tesla

A direct hit on a funding record is likely to rank in the top 10 results, depending on what they are. For example:

"results": [
        {
            "swirl_rank": 1,
            "swirl_score": 1316.565600582163,
            "searchprovider": "Company Funding Records (cloud/BigQuery)",
            "searchprovider_rank": 1,
            "title": "*Tesla* Motors",
            "url": "tesla-motors",
            "body": "*Tesla* Motors raised $40000000 series c on 2006-05-01. *Tesla* Motors is located in San Carlos CA and has 270 employees.",
            "date_published": "2006-05-01 00:00:00",
            "date_retrieved": "2023-01-11 12:16:43.302730",
            "author": "",
            "payload": {},
            "explain": {
                "stems": "tesla",
                "title": {
                    "tesla_*": 0.8357298742623626,
                    "Tesla_0": 0.8357298742623626,
                    "result_length_adjust": 4.5,
                    "query_length_adjust": 1.0
                },
                "body": {
                    "Tesla_0": 0.7187157993182859,
                    "result_length_adjust": 1.25,
                    "query_length_adjust": 1.0
                }
            }
        }

If a query begins with tag:, then only SearchProviders with that tag will be selected, regardless of their default status. (Of course, they must still be active.)

For example:

company: facebook

This will limit the query to SearchProviders with tag company.

SearchProviders do not need to have the default property set to true for Tags to work. So long as they have active set to true, then using the Tag in a query will cause Swirl to invoke it.

For more information, see Organizing SearchProviders with Active, Default and Tags

Relevancy Ranking

Swirl returns a unified result set consisting of results from all responding SearchProviders, matched by stemmed word form, and re-ranked using a cosine vector similarity relevancy model based on spaCy and normalized by query and token length. It also incorporates the original searchprovider_rank.

For more details please consult the Developer Guide Configure Relevancy Field Weights and Understand the Explain Structure sections.

Hit Highlighting

Swirl highlights occurrences of query terms in the title, body and author fields. For example:

    "body": "<em>Performance</em> <em>management</em> is the process of setting goals and expectations for employees and then tracking and measuring their progress. This can be done through regular one-on-one meetings, <em>performance</em> reviews, and other feedback mechanisms.",

As of version 1.10, Swirl can also integrate source synonym configurations into relevancy calculations with corresponding hit highlighting. See the Developer Guide Integrate Source Synonyms Into Swirl Relevancy section for details.

Using SearchProviders

SearchProviders are the essential element of Swirl. They make it quick and easy to search many sources - without writing any code.

SearchProviders are JSON objects. Swirl’s distribution comes preloaded with a variety of configurations for sources like Elastic, Solr, PostgreSQL, BigQuery, NLResearch.com, Miro.com, Atlassian, and more.

Swirl includes five (5) Google Programmable Search Engines (PSEs) to get you up and running right away. The credentials for these are shared with the Swirl Community.

SearchProvider Example JSON

• As of Release 1.10, Swirl includes example SearchProviders for JetBrains YouTrack Issues and Articles.

• As of Release 1.10.1, Swirl includes example SearchProviders for Atlassian Jira and Confluence Cloud products.

• As of Release 2.0, Swirl includes integration with Microsoft365 services.

• As of Release 2.0.3, Swirl includes support for RequestsPost with example SearchProvider JSON.

• As of Release 2.1, Swirl includes example SearchProviders for GitHub Code, Commits, Issues, and Pull Requests (public repositories only).

• As of Release 2.5, Swirl includes example SearchProviders for HubSpot Contacts, Companies, and Deals.
  • Requires creation of a “Private Apps” API token by the Super Admin with these scopes
  • In result_mappings, the url mapping must contain the HubSpot Org ID
• As of Release 2.5.1:
  • Swirl includes example SearchProviders for arXiv, EuropePMC, and LinkedIn profiles via a new Google PSE.
  • Both the ChatGPT Connector and QueryProcessor were updated to use OpenAI’s ChatCompletion method which supports the latest GPT models - including GPT-4 - and a much greater range of interactivity.
    • Three new Tags and one new query_mapping option are available for the ChatGPT SearchProvider to help shape the Prompt or Default Role passed to ChatGPT along with the user’s query.
    • The ChatGPT SearchProvider now queries the GPT-3.5-Turbo model by default.

• As of Release 2.6, Swirl includes SearchProviders for ServiceNow (Knowledge and Service Catalog), Google News and a searchable version of the Hacker News feeds.

• As of Release 3.0.0:
  • Swirl includes SearchProviders for Blockchain.com Bitcoin transactions and addresses as well as for Crunchbase organizations.
  • A new Google PSE SearchProvider that targets the new Swirl documentation website is included and enabled by default.
  • The EuropePMC SearchProvider is preloaded, set to active status, and configured to participate in Retrieval Augmented Generation (RAG) by default.

Activating

To activate a preloaded SearchProvider, edit it and change:

    "active": false

to

    "active": true

Click the PUT button to save the change.

Copy/Paste Install

If you have the raw JSON of SearchProvider, install it by copying/pasting into the form at the bottom of the SearchProvider endpoint.

1. Go to http://localhost:8000/swirl/searchproviders/
2. Click the Raw data tab on the form at the bottom of the page
3. Paste one SearchProvider’s JSON at a time into the form and press the POST button
4. Swirl will respond with the finished SearchProvider

Bulk Loading

Use the included swirl_load.py script to load any SearchProvider instantly, including lists of providers.

1. Open a terminal, cd into your <swirl-home> directory, and execute the following command:

   python swirl_load.py SearchProviders/provider-name.json -u admin -p your-admin-password
   

2. The script will load all SearchProvider configurations in the specified file at once and confirm.
3. Go to http://localhost:8000/swirl/searchproviders/ to see them!

Editing

Edit any SearchProvider by adding the id to the end of the /swirl/searchproviders URL.

For example: http://localhost:8000/swirl/searchproviders/1/

From here, you can use the form at the bottom of the page to:

• DELETE this SearchProvider, forever
• Edit the configuration of the SearchProvider and PUT the changes

Organizing SearchProviders with Active, Default and Tags

Three properties of SearchProviders are intended to allow expressive querying by targeting all or part of a query to groups of sources.

The suggestion is that SearchProviders who are good for most any search be left with Default set to True. Providers specific to a topic should have Default set to False and then "Tags": [ "topic1", "topic2" ] etc. When creating a search this ensures that the best providers for general querying are used when no searchprovider_list is specified. When the user desires to target a specific SearchProvider set, any combination of Tags or SearchProvider name and id values may be used freely.

Query Mappings

SearchProvider query_mappings are key/value pairs that define how to query a given SearchProvider.

They include field mappings and configurations that Swirl’s processors (like the AdaptiveQueryProcessor) use to align the query with each SearchProvider’s capabilities.

The following table summarizes the current query_mappings options:

Query Field Mappings

For query_mappings, keys that appear in the query_template wrapped in braces are replaced with the value.

    "url": "https://www.googleapis.com/customsearch/v1",
    "query_template": "{url}?cx={cx}&key={key}&q={query_string}",
    "query_processors": [
            "AdaptiveQueryProcessor"
        ],
    "query_mappings": "cx=0c38029ddd002c006,DATE_SORT=sort=date,PAGE=start=RESULT_INDEX",

At federation time, this becomes the following URL:

    https://www.googleapis.com/customsearch/v1?cx=0c38029ddd002c006&q=some_query_string

The url field is configured per SearchProvider. Any key/value pairs that never change for that SearchProvider should be put there.

The query_string is provided by Swirl as described in the Developer Guide

HTTP Request Headers

The optional http_request_headers field is available to all SearchProviders for passing custom HTTP Request Header values to the source alongside the user’s query. For example, the GitHub SearchProviders included in Swirl’s distribution use this feature to return a more accurate result snippet from GitHub that is then mapped to Swirl’s body field:

"http_request_headers": {
            "Accept": "application/vnd.github.text-match+json"
        },

"result_mappings": "title=name,body=text_matches[*].fragment, ...

Result Processors

In Release 2.5, important updates were made that affect the SearchProvider result_processors configuration.

1. Relevancy processing was split into two stages to improve performance
   • The revised CosineRelevancyPostResultProcessor must be added last in the Search.post_result_processors list.
   • Please review the JSON in the SearchProviders/ directory and update existing configurations to match.

"result_processors": [
            "MappingResultProcessor",
            "LenLimitingResultProcessor",
            "CosineRelevancyResultProcessor"
        ],

• This following of Error message in the Swirl logs indicates that one or more SearchProviders have not been updated:

INFO     search.py: invoking processor: CosineRelevancyPostResultProcessor
2023-07-31 16:31:39,268 ERROR    CosineRelevancyPostResultProcessor_2051: Error: Dictionary of result lengths is empty. Was CosineRelevancyResultProcessor included in Search Providers Processor configuration?

1. The DateFindingResultProcessor was added to the default Google PSE SearchProvider JSON. It finds a date in a large percentage of results that otherwise wouldn’t have one, and copies the date to the date_published field. Existing PSE SearchProvider configurations should be updated to include it:

"result_processors": [
            "MappingResultProcessor",
            "DateFinderResultProcessor",
            "CosineRelevancyResultProcessor"
        ],

Authentication & Credentials

The credentials property stores any required authentication information for the SearchProvider. The supported types are as follows:

key=value format

This credential is bound to the URL that is used to execute searches.

For example, from a Google PSE:

    "credentials": "key=your-google-api-key-here",
    "query_template": "{url}?cx={cx}&key={key}&q={query_string}",

bearer=token format

Bearer tokens are supported by the RequestsGet and RequestsPost connectors. They are sent with the request header.

For example, from the Miro SearchProvider:

    "credentials": "bearer=your-miro-api-token",

X-Api-Key=key format

X-Api-Keys are supported by the RequestsGet and RequestsPost connectors. They are sent with the request header.

    "credentials": "X-Api-Key=<your-api-key>",

HTTPBasicAuth, HTTPDigestAuth, HTTPProxyAuth

These methods are supported by the RequestsGet, ElasticSearch and OpenSearch connectors.

For example, from the Solr with Auth SearchProvider:

    "credentials": "HTTPBasicAuth('solr-username','solr-password')",

Other Credentials

Consult the Developer Guide for details on how to Develop New Connector.

Response Mappings

SearchProvider response_mappings determine how each source’s response is normalized into JSON. They are used by the Connector normalize_response method. Each mapping is a JSONPath.

Here is the response_mappings from a Google PSE:

    "response_mappings": "FOUND=searchInformation.totalResults,RETRIEVED=queries.request[0].count,RESULTS=items",

The following table summarizes the response_mappings options:

Result Mappings

SearchProvider result_mappings determine how each source result set, in JSON format, is mapped to the Swirl result schema. Each mapping is a JSONPath.

Here is a result_mapping from a Google PSE:

    "result_mappings": "url=link,body=snippet,author=displayLink,cacheId,pagemap.metatags[*].['og:type'],pagemap.metatags[*].['og:site_name'],pagemap.metatags[*].['og:description'],NO_PAYLOAD",

The mappings url=link and body=snippet map the Swirl result fields to the corresponding Google PSE result fields.

For Release 2.5.1, requests.py was updated to handle XML responses from source APIs and convert them to JSON for mapping in SearchProvider configurations.

Multiple Mappings

As of version 1.6, Swirl can map multiple SearchProvider fields to a single Swirl field, aggregating multiple responses in the PAYLOAD field as necessary.

For example:

"result_mappings": "body=content|description,..."

If only one field, content or description, are populated for a response, then that will be mapped to the Swirl’s body field. Should both fields be populated, the second field is moved to the PAYLOAD and named <swirl-field>_<source_field>. For example:

        {
            "swirl_rank": 1,
            "swirl_score": 24391.4814426326,
            "searchprovider": "Latest News (web/newsdata.io)",
            "searchprovider_rank": 2,
            "title": "What The *Mid-Term* *Elections* Mean For U.S. Energy",
            "url": "https://www.forbes.com/sites/davidblackmon/2022/11/13/what-the-mid-term-elections-mean-for-us-energy/",
            "body": "Leaders in U.S. domestic energy sectors should expect President Joe Biden to feel emboldened in the wake of *mid-term* *elections* that saw voters issue a status quo verdict on his policies.",
            "date_published": "2022-11-13 13:38:30",
            "date_retrieved": "2022-11-13 18:03:33.676767",
            "author": "David Blackmon",
            "payload": {
                "body_description": "Leaders in U.S. domestic energy sectors should expect President Joe Biden to feel emboldened in the wake of mid-term elections that saw voters issue a status quo verdict on his policies.",
                "source_id": "forbes",
                "category": [
                    "business"
                ]

Result Mapping Options

The following table explains the result_mappings options:

BLOCK Example

For a working example of the new BLOCK configuration, check out the Swirl 2.0 default ChatGPT configuration with the ai_summary result block:

{
    "name": "ChatGPT - OpenAI",
    "active": false,
    "default": true,
    "connector": "ChatGPT",
    "url": "",
    "query_template": "",
    "query_processors": [
        "AdaptiveQueryProcessor"
    ],
    "query_mappings": "PROMPT='Tell me about: {query_to_provider}'",
    "result_processors": [
        "GenericResultProcessor",
        "CosineRelevancyResultProcessor"
    ],
    "response_mappings": "",
    "result_mappings": "BLOCK=ai_summary",
    "results_per_query": 10,
    "credentials": "your-openai-API-key-here",
    "tags": [
        "ChatGPT",
        "Question"
    ]
}

You must remove the default result_mappings value of BLOCK=ai_summary in the SearchProvider configuration to enable the ChatGPT or Question Tags! Otherwise, these Tags will be ignored.

The configuration is noted in the appropriate info blocks of Results as well:

"info": {
        ...
        "ChatGPT (web/OpenAI)": {
            ...
            "result_block": "ai_summary",
            ...
        }
        ...
        "results": {
            ...
            "result_blocks": [
                "ai_summary"
            ],
            ...
        }
},
"ai_summary": [
        {
            ...
            "searchprovider": "ChatGPT (web/OpenAI)",
            ...

Date Published Display

As of version 2.1, a date_published_display option is available, allowing the mapping of different values for date_published and date_published_display in SearchProvider configurations:

"result_mappings":  ... date_published=foo.bar.date1,date_published_display=foo.bar.date2 ...

Results objects will reflect both date values:

"date_published": "2010-01-01 00:00:00",
"date_published_display": "c2010",

Result Schema

The json_result schema for each result in the Result list is defined by the create_result_dictionary() function in swirl/processors/utils.py.

Result mixers further manipulate and re-organize the data from multiple results.

The Result schema can be seen in swirl/models.py

PAYLOAD Field

The PAYLOAD is a JSON list structure that can hold arbitrary data structures. This is only meaningful if you also specify NO_PAYLOAD. If you don’t specify this keyword, you will get all of the SearchProvider’s raw data mapped into this field.

After mapping the fields you want the way you want them, then add this directive to the result_mappings so that you only get back what you want.

To use NO_PAYLOAD most effectively, send your first query to a SearchProvider without it to see what you get back in the PAYLOAD.