LogoLogo
👋 Get in touch⭐️ GitHub
  • Welcome
  • Getting Started
    • Why Superlinked?
    • Setup Superlinked
    • Basic Building Blocks
  • Run in Production
    • Overview
    • Setup Superlinked Server
      • Configuring your app
      • Interacting with app via API
    • Supported Vector Databases
      • Redis
      • Mongo DB
      • Qdrant
  • Concepts
    • Overview
    • Combining Multiple Embeddings for Better Retrieval Outcomes
    • Dynamic Parameters/Query Time weights
  • Reference
    • Overview
    • Changelog
    • Components
      • Schema
        • Id Schema Object
        • Event Schema
        • Event Schema Object
        • Schema Object
        • Schema
      • Parser
        • Json Parser
        • Dataframe Parser
        • Data Parser
      • Dag
        • Period Time
      • Storage
        • Vector Database
        • Qdrant Vector Database
        • Mongo Db Vector Database
        • Redis Vector Database
        • In Memory Vector Database
      • Space
        • Custom Space
        • Space Field Set
        • Input Aggregation Mode
        • Text Similarity Space
        • Space
        • Categorical Similarity Space
        • Recency Space
        • Number Space
        • Exception
        • Has Space Field Set
        • Image Space Field Set
        • Image Space
      • Query
        • Query Mixin
        • Query Param Value Setter
        • Query Weighting
        • Space Weight Param Info
        • Clause Params
        • Query Descriptor
        • Query
        • Param Evaluator
        • Param
        • Result
        • Query Filter Information
        • Query Filter Validator
        • Natural Language Query Param Handler
        • Query Filters
        • Query Param Information
        • Nlq Param Evaluator
        • Query Vector Factory
        • Nlq Pydantic Model Builder
        • Typed Param
        • Query Clause
        • Nlq
          • Nlq Compatible Clause Handler
          • Nlq Handler
          • Nlq Clause Collector
          • Exception
          • Suggestion
            • Query Suggestions Prompt Builder
            • Query Suggestion Model
          • Param Filler
            • Query Param Model Validator Info
            • Nlq Annotation
            • Query Param Model Builder
            • Query Param Prompt Builder
            • Query Param Model Validator
            • Templates
        • Query Clause
          • Hard Filter Clause
          • Space Weight Map
          • Looks Like Filter Clause
          • Similar Filter Clause
          • Base Looks Like Filter Clause
          • Single Value Param Query Clause
          • Radius Clause
          • Select Clause
          • Overriden Now Clause
          • Nlq System Prompt Clause
          • Looks Like Filter Clause Weights By Space
          • Limit Clause
          • Weight By Space Clause
          • Nlq Clause
          • Query Clause
        • Predicate
          • Binary Op
          • Binary Predicate
          • Query Predicate
        • Query Result Converter
          • Default Query Result Converter
          • Query Result Converter
          • Serializable Query Result Converter
      • Executor
        • Executor
        • Exception
        • Query
          • Query Executor
        • Rest
          • Rest Handler
          • Rest Configuration
          • Rest Descriptor
          • Rest Executor
        • Interactive
          • Interactive Executor
        • In Memory
          • In Memory Executor
      • App
        • App
        • Online
          • Online App
        • Rest
          • Rest App
        • Interactive
          • Interactive App
        • In Memory
          • In Memory App
      • Source
        • Interactive Source
        • Data Loader Source
        • In Memory Source
        • Source
        • Rest Source
        • Types
      • Index
        • Index
        • Effect
        • Util
          • Event Aggregation Effect Group
          • Aggregation Effect Group
          • Aggregation Node Util
          • Effect With Referenced Schema Object
          • Event Aggregation Node Util
      • Registry
        • Superlinked Registry
        • Exception
  • Recipes
    • Overview
    • Multi-Modal Semantic Search
      • Hotel Search
    • Recommendation System
      • E-Commerce RecSys
  • Tutorials
    • Overview
    • Semantic Search - News
    • Semantic Search - Movies
    • Semantic Search - Product Images & Descriptions
    • RecSys - Ecommerce
    • RAG - HR
    • Analytics - User Acquisition
    • Analytics - Keyword Expansion
  • Help & FAQ
    • Logging
    • Support
    • Discussion
  • Policies
    • Terms of Use
    • Privacy Policy
Powered by GitBook
On this page

Was this helpful?

Edit on GitHub
  1. Tutorials

RAG - HR

Learn about fundamental concepts of Superlinked.

Many companies aim to make HR documents more accessible to employees and contractors. A recent promising approach is building a RAG system on top of HR documents. This is because it

  • provides a unique response to any query,

  • reduces hallucinations as the answer has to be grounded in the retrieved context,

  • makes the process highly scalable due to automatic question-answering.

However, this solution possesses its unique challenges. For example

  • keeping the knowledgebase consistent and up-to-date,

  • running LLMs efficiently,

  • ensuring the generated results are correct and aligned with company communication guidelines,

In our case, we have three HR policy sources for our hypothetical company.

  1. An older HR policy, which contains our maternity leave policy and details on manager responsibilities.

  2. A more recent update, that contains inaccurate information on management responsibilities - but one that also contains unique information on paternity leave.

  3. A newer policy document that contains updated information about management responsibilities, correcting the mistakes of the previous modernising attempt - alongside some other HR policy information.

A good system will be able to:

  • provide relevant information on maternity leave (only covered in the old document),

  • synthesize contradicting information and only present to us the correct ones

Regarding synthesizing information, there are contradictions between the documents pertaining to management responsibilities. We are going to use the

  • creation_date of the policy,

  • and the usefulness score of each paragraph

as a proxy to know for similar documents with

  • similar information but different wording, and some important parts slightly altered, or

  • also talking about seemingly the same topic, but without conveying useful information

which one is more relevant to our query.

Follow along in this Colab

PreviousRecSys - EcommerceNextAnalytics - User Acquisition

Last updated 7 months ago

Was this helpful?

LogoGoogle Colab