Table of Contents


NeuralSeek is an AI-powered answer generation engine. It works by taking a user question and assembling a courpus of backing information from a KnowledgeBase like ElasticSearch or Watson Discovery.

NeuralSeek uses this corpus to conduct just-in-time training to generate a conversational answer to the user question. As a user you just see an answer in natural language – not options, paragraphs or references from source material. NeuralSeek attempts to fill in the gaps and join related thoughts – just like a live customer support agent would do in trying to answer user questions. NeuralSeek works independant of any virtual agent or KnowledgeBase – you can connect it to anything that can call an api or webhook. NeuralSeek supports no-code connections to several ChatBot / Virtual Agent and Knowledge Base systems.


Provision NeuralSeek at
After Provisioning, click on the “Launch NeuralSeek” button. You’ll land on the “Configure” tab. Input your KnowledgeBase Details, and set the Company name. Click “Save” on the bottom of the page. NeuralSeek will test the connection.

Click on the “Seek” Tab, and ask a question or two! The seek tab allows you to test questions and view responses and confidence levels.

Responses can be provided in English, Spanish, Portuguese, French, German, Italian, Arabic, and Japanese. Questions do not need to be asked in the same language as the requested response. EG: You can ask a question in French and request the output in Spanish.

Integrate with Watson Assistant. Follow the Directions on the “Integrate” tab to hook NeuralSeek up to your assistant in a few easy clicks.

Automatic Data Cleansing & Prep

NeuralSeek will automatically cleanse data inside of the KnowledgeBase it is connected to, depending on the content type and permissions given.
NeuralSeek will automatically cleanse scraped webpages inside your KnowledgeBase to remove nuisance text such as banners and cookie warnings – in order to expose more of your company’s relevant information for real-time use. It is recommended that your web scrape is set to update at a weekly or monthly rate. If using Watson Discovery, hourly and daily updates will significantly consume Watson Discovery “query” allocation, as a query is used for each document that is cleansed.

KnowledgeBase Tips

We automatically apply enhancements to the project that you connect to NeuralSeek.

Avoid conflicting information! If you upload PDF or other corporate docs, be sure to remove old docs that may conflict with the new ones.

Feedback Helps! Any time spent curating or training results inside the KnowledgeBase will help expose preferred answers to questions inside NeuralSeek, however we aim to provide a solid baseline with our automated enhancements so that many questions may be answered sufficiently without training.

Integrate with Watson Assistant

When consuming NeuralSeek from inside Watson Assistant, most integrations will want to use the “Actions” based extensions. Depending on the response time of your corporate KnowledgeBase, NeuralSeek may sometimes take longer to return than the 7 second timeout of Watson Assistant imposes on webhooks. Lower tiers of Watson Discovery have the longest response time. Higher tiers of Watson Discovery, ElasticSearch, and NeuralSeek’s own fully-integrated Small-Business plan provide for faster responses.

You can still use the custom extension if using Dialogs. Simply hook an “anything_else” dialog node to push the user query to an action that calls the extension.

The “Integrate” tab walks you through the steps to install the custom extension.

You can also use NeuralSeek with dialog-based Watson Assistant. Set up an action with the NeuralSeek Extension, and then call that from an “anything_else” node in Dialogs. It is highly recommended that if you take this approach that you manually “null” the returned context variable after you display the response text. If you do not, in certain situations your assistant will max out the allowed context size and stop responding to user questions until a new session is started.

NeuralSeek REST API

NeuralSeek can be connected and consumed by any chatbot, Virtual Agent, webpage, or system that can make an external API call or consume a webhook, as NeuralSeek provides an open API endpoint. You can find the details of the NeuralSeek API at

Conversational Context

NeuralSeek maintains conversational context during interaction with a user. A session token is required to be sent with all requests to enable context features – this lets us track the conversation across multiple independent calls. NeuralSeek employs a multitude of NLP models to break down the structure of both questions and their generated responses. This enables NeuralSeek to estimate the topic of the conversation and keep the interaction focused in follow-on questions that do not specifically mention the primary subject. In addition, these NLP models enable NeuralSeek to filter corporate knowledge topically by date to ensure that the information being returned is focused on the time period of the question.

Integrate with a corporate KnowledgeBase

NeuralSeek currently supports no-code integrations to both IBM Watson Discovery and Elastic App Search. No other inbound integrations are currently supported. If you do not have an existing compatible corporate KnowledgeBase, NeuralSeek’s Small-Business plan is pre-integrated with one and offers 30% faster response time as compared to bringing your own corporate KnowledgeBase.

Language Support

NeuralSeek can answer and respond in multiple languages. NeuralSeek currently supports taking questions and delivering answers in English, Spanish, Portuguese, French, German, Italian, Arabic, Korean, Chinese, and  Japanese. To switch to another language, select the language from the dropdown in the NeuralSeek UI, or set the language code dynamically as you call the Seek endpoint.

Context keeping and other advanced features are only supported in English
Languages will perform best when both the knowledge repository articles and the user question are in the selected output language. EG: A Spanish question will work best with output set to Spanish and the knowledge repository loaded and trained with Spanish-language articles.

Output may be inconsistent with user queries and knowledge base context given in a language other than the requested output language. EG: if the output language is set to English, but the user asks a question in German against a German knowledge base – you may get an answer either in German or English.


The quality of the output of NeuralSeek is directly correlated to the quality of the knowledge base documents loaded into Watson Discovery. If you are getting undesired output from NeuralSeek, investigate the content that you have provided in Discovery. The more the better. Ensure the content is not contradictory.

NeuralSeek is designed to always give an answer. This means that NeuralSeek will sometimes give a wrong answer. This is an unavoidable consequence of a system designed to always answer. You can use the Warning & Minimum confidence settings in NeuralSeek to block the system from answering or provide a warning in low confidence instances.

Curation of answers to a Virtual Agent

NeuralSeek will automatically generate Watson Assistant “Actions” or “Dialogs”, based on user questions that are asked. Generally Watson Assistant needs 5 or more user question examples to train on for a high-confidence match to a user query. When user questions are cataloged by the system, NeuralSeek automatically tries to generate similar worded questions to meet the minimum of 5 user examples. Similar Question generation may take up to 1 minute to show inside the Curate tab after a new user question is logged.

Select Intents on the curate tab that you want to curate into Watson Assistant. NeuralSeek will highlight other intents that you should NOT select because they will share training data and collide. NeuralSeek attempts to group like questions by keywords, so it is expected that the same questions and answers may be in multiple intents. You need to choose the grouping that works best for your use case. Once you download your Actions file, upload it into Watson Assistant by using the gear icon in the upper right of the Actions screen. If you already have actions, it is important to first download them and upload them as “Base Actions” into NeuralSeek. NeuralSeek will then preserve all of the existing actions as you export your new actions.

Dynamic Personalization

NeuralSeek supports creating personalized answers. This can be previewed in the “Seek” tab, and in production environments you will pass the personalization details via our API as the REST call to /seek is made. When using personalization, the conversation shifts from third person to first person. NeuralSeek will try to work the details fed into the personalization options into the response, where they make most natural sense.  If many details are given, not every detail will be returned in every response in order to make the conversation feel more natural.

Personalized answers are not eligible for curation, and are not viewable in the curation logs.

Round Trip Logging

NeuralSeek supports receiving logs from Virtual Agents in order to monitor curated responses. For Watson Assistant, follow the setup instructions on the Integrate tab of the NeuralSeek UI. Once enabled, NeuralSeek will star any intent that has been touched in the Virtual Agent, and will monitor those curated responses to identify if the source documentation for those answers has been updated. If the answer is found to be out of date, NeuralSeek will notify you by adding an icon to the intent in the Curate page. You may click on that icon to automatically regenerate answers based on the updated documentation.


NeuralSeek monitors questions and generated answers for patterns, and will create analytics for you automatically about the questions your users are asking about your documentation. The Analytics tab in the UI holds this data. At the top of the page there is an executive summary, highlighting the areas NeuralSeek feels need the most attention, based on frequency of the questions and NeuralSeek’s coverage and confidence scores.

Coverage Score:  Coverage is a measure of how many documents or sections of a document talk about the subject area of a user question. A low coverage score is not necessarily bad, depending on the question. A high coverage score, on the other hand may be indicative of questions that have conflicting or confusing source material. You should look to ensure your KnowledgeBase does not have contradictory information in it.

Confidence Score:  How well does NeuralSeek believe that the corporate knowledge found answers the user question. The higher score here, the better.