Vectara Answer is a sample app for Vectara-powered Summarized Semantic Search (or question-answering) with advanced configuration options. For examples of what you can build with Vectara Answer, check out Ask News, LegalAid, or any of the other demo applications.
Note: For more user interface options and a simpler codebase, use Create-UI.
Tip
Looking for something else? Try another open-source project:
- React-Chatbot: Add a compact Vectara-powered chatbot widget chat to your React apps.
- Create-UI: The fastest way to generate a working React codebase for a range of generative and semantic search UIs.
- Vectara Ingest: Configurable crawlers for pulling data from many popular data sources and indexing into Vectara.
To get started, the minimum requirement is to install npm and node. That's it!
Vectara Answer comes packaged with preset configurations that allow you to spin up a sample application using Vectara's public datastores. To quickly get started, run the following command:
npm run bootstrap
When prompted for which application to create, select from one of three default apps:
Vectara Docs- Answer questions about Vectara documentationVectara.com- Answer questions about the content of the Vectara company websiteAskFeynman- Answer questions about Richard Feynman's lectures
After selecting which application to create, you'll have the app running in your browser at http://localhost:4444.
Congratulations! You've just setup and run a sample app powered by Vectara!
The bootstrap command installs dependencies, runs the configuration script to generate an .env file
(which includes all the needed configuration parameters), and spins up the local application.
If you would like to run the setup steps individually, you can run:
npm install: for installing dependenciesnpm run configure: for running the configuration scriptnpm run start: for running the application locally
When building your own application, you will need to:
- Create a data store: Log into the Vectara Console and create a corpus.
- Add data to the corpus: You can use Vectara Ingest to crawl datasets and websites, upload files in the Vectara Console, or use our Indexing APIs directly.
Tip
When ingesting data into your corpus you are free to add metadata to any indexed document. By convention, if you add the url field as metadata, Vectara Answer will use the value in that field to display a clickable link for each reference or citations it display on the results page.
When running npm run bootstrap, if you choose [Create Your Own] from the application selection prompt, you will be asked to provide:
- Your Vectara customer ID
- The ID of the corpus you created
- The API key of your selected Vectara corpus (NOTE: Depending on your set up, this may be visible to users. To ensure safe sharing, ensure that this key is set up to only have query access.)
- Any sample questions to display on the site, to get your users started.
Once provided, the values above will go into your own customized configuration (.env file), and your site will be ready to go via npm start.
By default, Vectara Answer runs locally on your machine using npm run start. There is also an option to use Vectara Answer with Docker, which also makes it easy to deploy Vectara Answer to a cloud environment.
Please see these detailed instructions for more details on using Docker.
After the configuration process has created your .env file (as part of the bootstrap process), you are free to make modifications to it to suit your development needs.
Note that the variables in the .env file all have the REACT_APP prefix, as is needed to be recognized by Vectara Answer code.
# These config vars are required for connecting to your Vectara data and issuing requests.
customer_id: 123456789
corpus_id: 5
corpus_key: vectara_docs_1
api_key: "zqt_abcdef..."Notes:
- Vectara APIV2 uses
corpus_key, but the oldercorpus_idis also supported for backwards compatibility. We encourage you to usecorpus_keyas we may deprecate support for V1 in the future. corpus_key(orcorpus_id) can be a set of corpora in which case each query runs against all those corpora. In such a case, the format is a comma-separated list of corpus keys or corpus IDs, for example:
corpus_id: "123,234,345"OR
corpus_key: "vectara_docs_1,vectara_website_3"These configuration parameters enable you to configure the look and feel of the search header. The search header may include a logo, a title (text) and a description. Most commonly we just have to define a title and description.
Alternatively, the title can be disabled and instead you can use a logo instead, but note that currently the logo.png file must be added to the codebase under the config_images folder.
# Define the title of your application (renders next to the logo, if exists).
search_title: "My application"
# Define the description to render opposite the logo and title.
search_description: "Data that speaks for itself"
# Define the placeholder text inside the search box.
search_placeholder: "Ask me anything"
# Define the URL the browser will redirect to when the user clicks the logo above the search controls.
search_logo_link: "https://asknews.demo.vectara.com"
# Define the logo that appears in the search header. Any images you place in your `config_images` directory will be available.
search_logo_src: "config_images/logo.png"
# Describe the logo for improved accessibility.
search_logo_alt: "Vectara logo"
# Customize the height at which to render the logo. The width will scale proportionately.
search_logo_height: 20The way summarization works can be configured as follows:
# Switches the mode of the ux to "summary" mode or "search" mode (if not specified defaults to "summary" mode). When set to "summary", a summary is shown along with references (citations) used in the summary. When set to "search", only search results are shown and no call is made to the summarization API.
ux: "summary"
# Default language for summary response (if not specified defaults to "auto").
# accepts two-letter code (en) or three-letter code (eng)
summary_default_language: "eng"
# Number of sentences before and after relevant text segment used for summarization.
summary_num_sentences: 3
# Number of results used for summarization.
summary_num_results: 10
# The name of the summarization prompt in Vectara.
# If unspecified, the default for the account is used.
# See https://docs.vectara.com/docs/learn/grounded-generation/select-a-summarizer for available summarization prompts. Note that some prompts are only available to Vectara scale customers.
summary_prompt_name: vectara-summary-ext-v1.2.0
# Filename for a local file (same folder as your config.yaml file) which includes the text for a custom prompt
# If specified, used to replace the prompt text used by the prompt_name. You can see an example prompt.txt file in the config/ask-feynman folder.
# Note this is a Scale only feature and is not yet enabled for streaming.
summary_prompt_text_filename: prompt.txt
# Whether to disable or enable factual consistency score as (score or badge) based on the HHEMv2 (based on https://huggingface.co/vectara/hallucination_evaluation_model).
# default value is disable. To enable it set it to score or badge.
summary_fcs_mode: score
# Add multiple prompts option for the summary.
summary_prompt_options: vectara-summary-ext-24-05-med-omni,vectara-summary-ext-24-05-sml
To stream the results, set the following parameter to True.
# Streaming on/off for summarization
enable_stream_query: TrueHybrid search is a capability that combines the strength of neural (semantic) search with traditional keywords search. By default, Vectara Answer utilizes hybrid search with lambda=0.1 for short queries (num_words<=2) and lambda=0.0 (pure neural search) otherwise, but you can define other values here.
# hybrid search
hybrid_search_num_words: 2
hybrid_search_lambda_long: 0.0
hybrid_search_lambda_short: 0.1Vectara Answer can display an application header and footer for branding purposes. These configuration parameters allow you to configure the look and feel of this header and footer.
# Hide or show the app header.
enable_app_header: False
# Hide or show the app footer.
enable_app_footer: False
# Define the title of your app to render in the browser tab.
app_title: "Your application title here"
# Define the URL the browser will redirect to when the user clicks the logo in the app header.
app_header_logo_link: "https://www.vectara.com"
# Define the logo that appears in the app header. Any images you place in your `config_images` directory will be available.
app_header_logo_src: "config_images/logo.png"
# Describe the logo for improved accessibility.
app_header_logo_alt: "Vectara logo"
# Customize the height at which to render the logo. The width will scale proportionately.
app_header_logo_height: 20If your application uses more than one corpus, you can define source filters to enable the user to narrow their search to a specific corpus. This feature assumes the following:
- You have defined a
sourcemeta-data field on the Vectara corpus - During data ingestion, you've added the source text to each document appropriately (in the
sourcemetadata field)
The following parameters control how the sources feature works:
# Hide or show source filters.
enable_source_filters: True
# whether the "all source" button should be enabled or not (default True)
all_sources: True
# A comma-separated list of the sources on which users can filter.
sources: "BBC,NPR,FOX,CNBC,CNN"The sources parameters is a comma-separated list of source names that will be displayed underneath the search bar, and the user can select
if results returned should be from "all source" or one of the selected sources.
In this case you must specify corpus_id (see above) to be the list of matching corpus IDs, also comma separated.
For example:
enable_source_filters: True
all_sources: True
sources: "BBC,NPR,FOX,CNBC,CNN"
corpus_id: "123,124,125,126,127"If all_sources is set to False, the application will only display the individual source but not the "All sources" button.
This means the user will only be able to select a specific source for each query.
Whether to use Vectara's reranking functionality. Note that reranking currently works for English language only, so if the documents in your corpus are in other languages, it's recommended to set this to "False".
# choose a reranker from the following choices, You can use single or multiple rerankers with coma separated. e.g `slingshot,mmr`
# if you are using multiple reranker please make sure slingshot should be first if you are using it. For more details [chain reranking](https://docs.vectara.com/docs/learn/chain-reranker)
reranker_name: normal | slingshot | mmr | userfn | slingshot,mmr,userfn
# number of results to use for reranking
rerank_num_results: 50
# enables users to define custom reranking functions using document-level metadata, part-level metadata.
user_function: "if (now() < iso_datetime_parse('2024-12-04T10:14:50Z')) 1 else 2" To use Vectara's MMR (Maximum Marginal Relevance) functionality please set the reranker_name = mmr, and add a mmr_diversity_bias value.
# Diversity bias factor (0..1) for MMR reranker. The higher the value, the more MMR is preferred over relevance.
mmr_diversity_bias: 0.3vectara-answer supports Google SSO front-end authentication. Note this is not a fully secure SSO solution, and you should consult your security experts for in a full enterprise deployment.
# Configure your app to require the user to log in with Google SSO.
authenticate: True
google_client_id: "cb67dbce87wcc"# Track user interaction with your app using Google Analytics.
google_analytics_tracking_code: "123456789"
# Track user interaction with your app using Google Tag Manager.
gtm_container_id: "GTM-1234567"Customize the result display, this provides options to tailor the presentation of the results shown.
# Enable the "explore related content" link, which does a search for content related to the text in each results's matching snippet
related_content: True# Track user experience with Full Story
full_story_org_id: "org1123"In addition to customization via configuration, you can customize Vectara Answer further by modifying its code directly in your own fork of the repository.
The UI source code is all in the src/ directory. See the UI README.md to learn how to make changes to the UI source.
While npm run start runs the application with a local client that accesses the Vectara API directly, running the app via Docker (see below) spins up a full-stack solution, using a proxy server to make Vectara API requests.
To modify the request handlers, make changes to /server/index.js.
👤 Vectara
- Website: https://vectara.com
- Twitter: @vectara
- GitHub: @vectara
- LinkedIn: @vectara
- Discord: @vectara
Contributions, issues and feature requests are welcome!
Feel free to check issues page. You can also take a look at the contributing guide.
Give a ⭐️ if this project helped you!
Copyright © 2024 Vectara.
This project is Apache 2.0 licensed.