Langchain/LlamaIndex provide easy to use abstractions that can be used for quick experimentation and prototyping on jupyter notebooks. But, when things move to production, there are constraints like the components should be modular, easily scalable and extendable. This is where Cognita comes in action. Cognita uses Langchain/Llamaindex under the hood and provides an organisation to your codebase, where each of the RAG component is modular, API driven and easily extendible. Cognita can be used easily in a local setup, at the same time, offers you a production ready environment along with no-code UI support. Cognita also supports incremental indexing by default.

You can try out Cognita at: https://cognita.truefoundry.com

• [June, 2024] Added one click local deployment of cognita. You can now run the entire cognita system using docker-compose. This makes it easier to test and develop locally.
• [May, 2024] Added support for Embedding and Reranking using Infninty Server. You can now use hosted services for variatey embeddings and reranking services available on huggingface. This reduces the burden on the main cognita system and makes it more scalable.
• [May, 2024] Cleaned up requirements for optional package installations for vector dbs, parsers, embedders, and rerankers.
• [May, 2024] Conditional docker builds with arguments for optional package installations
• [April, 2024] Support for multi-modal vision parser using GPT-4

• Cognita
  • Why use Cognita?
• 🎉 What's new in Cognita
• Contents
  • Introduction
    • Advantages of using Cognita are:
    • Features:
• 🚀 Quickstart: Running Cognita Locally
  • 🐳 Using Docker compose (recommended)
  • Cognita from source
• ⚒️ Project Architecture
  • Cognita Components:
  • Data Indexing:
  • ❓ Question-Answering using API Server:
  • 💻 Code Structure:
  • Customizing the Code for your usecase
    • Customizing Dataloaders:
    • Customizing Embedder:
    • Customizing Parsers:
    • Adding Custom VectorDB:
    • Rerankers:
• 💡 Writing your Query Controller (QnA):
  • Steps to add your custom Query Controller:
• 🐳 Quickstart: Deployment with Truefoundry:
  • Using the RAG UI:
• 💖 Open Source Contribution
• 🔮 Future developments
• Star History

Cognita is an open-source framework to organize your RAG codebase along with a frontend to play around with different RAG customizations. It provides a simple way to organize your codebase so that it becomes easy to test it locally while also being able to deploy it in a production ready environment. The key issues that arise while productionizing RAG system from a Jupyter Notebook are:

1. Chunking and Embedding Job: The chunking and embedding code usually needs to be abstracted out and deployed as a job. Sometimes the job will need to run on a schedule or be triggered via an event to keep the data updated.
2. Query Service: The code that generates the answer from the query needs to be wrapped up in a api server like FastAPI and should be deployed as a service. This service should be able to handle multiple queries at the same time and also autoscale with higher traffic.
3. LLM / Embedding Model Deployment: Often times, if we are using open-source models, we load the model in the Jupyter notebook. This will need to be hosted as a separate service in production and model will need to be called as an API.
4. Vector DB deployment: Most testing happens on vector DBs in memory or on disk. However, in production, the DBs need to be deployed in a more scalable and reliable way.

Cognita makes it really easy to customize and experiment everything about a RAG system and still be able to deploy it in a good way. It also ships with a UI that makes it easier to try out different RAG configurations and see the results in real time. You can use it locally or with/without using any Truefoundry components. However, using Truefoundry components makes it easier to test different models and deploy the system in a scalable way. Cognita allows you to host multiple RAG systems using one app.

1. A central reusable repository of parsers, loaders, embedders and retrievers.
2. Ability for non-technical users to play with UI - Upload documents and perform QnA using modules built by the development team.
3. Fully API driven - which allows integration with other systems.

   If you use Cognita with Truefoundry AI Gateway, you can get logging, metrics and feedback mechanism for your user queries.

1. Support for multiple document retrievers that use Similarity Search, Query Decompostion, Document Reranking, etc
2. Support for SOTA OpenSource embeddings and reranking from mixedbread-ai
3. Support for using LLMs using Ollama
4. Support for incremental indexing that ingests entire documents in batches (reduces compute burden), keeps track of already indexed documents and prevents re-indexing of those docs.

Cognita and all of it's services can be run using docker-compose. This is the recommended way to run Cognita locally. You can run the following command to start the services:

docker-compose --env-file compose.env up --build

• The compose file uses compose.env file for environment variables. You can modify it as per your needs.
• The compose file will start the following services:
  • ollama-server - Used to start local LLM server. compose.env has OLLAMA_MODEL as the environment variable to specify the model.
  • infinity-server - Used to start local embeddings and rerankers server. compose.env has INFINITY_EMBEDDING_MODEL and INFINITY_RERANKING_MODEL as environment variables to specify the embedding and reranker from HuggingFace Hub.
  • qdrant-server - Used to start local vector db server.
  • cognita-backend - Used to start the FastAPI backend server for Cognita.
  • cognita-frontend - Used to start the frontend for Cognita.
• Once the services are up, you can access the infinity server at http://localhost:7997, qdrant server at http://localhost:6333, the backend at http://localhost:8000 and frontend at http://localhost:5001.
• Backend uses local.metadata.yaml file for configuration. You can modify it as per your needs. The file is used to setup collection name, different data source path, and embedder configurations. Before starting of backend an indexer job is run to index the data sources mentioned in local.metadata.yaml file.

⚠️ Note: Currently UI supports only QnA and not the data source and collection creation. These have to be done via local.metadata.yaml only. Post that restart the docker-compose services. The work is in progress to bring that facility via UI as well and make the experience seamless.

You can play around with the code locally using the python script or using the UI component that ships with the code.

Before you can use Cognita, you'll need to ensure that Python >=3.10.0 is installed on your system and that you can create a virtual environment for a safer and cleaner project setup.

It's recommended to use a virtual environment to avoid conflicts with other projects or system-wide Python packages.

Navigate to your project's directory in the terminal. Run the following command to create a virtual environment named venv (you can name it anything you like):

python3 -m venv ./venv

• On Windows, activate the virtual environment by running:

venv\Scripts\activate.bat

• On macOS and Linux, activate it with:

source venv/bin/activate

Once your virtual environment is activated, you'll see its name in the terminal prompt. Now you're ready to install Cognita using the steps provided in the Quickstart sections.

Remember to deactivate the virtual environment when you're done working with Cognita by simply running deactivate in the terminal.

Following are the instructions for running Cognita locally without any additional Truefoundry dependencies

In the project root execute the following command:

pip install -r backend/requirements.txt

• Install packages for additional parsers like PDFTableParser that uses deep doctection for table extraction from PDFs. This is optional and can be skipped if you don't need to extract tables from PDFs.

  pip install -r backend/parsers.requirements.txt
  

• Install packages for vector_db that uses singlestore for vector db. This is optional and can be skipped if you don't need to use singlestore.

  pip install -r backend/vectordb.requirements.txt
  

  Uncomment the respective vector db in backend/modules/vector_db/__init__.py to use it.

• Rerankers and Embedders are to be used via hosted services like Infinity. Respective service files can be found under embedder and reranker directories.

• To install Infinity service, follow the instructions here. You can also run the following command to start a Docker container having mixedbread embeddings and rerankers.

    docker run -it --gpus all \
    -v $PWD/infinity/data:/app/.cache \
    -p 7997:7997 \
    michaelf34/infinity:latest \
    v2 \
    --model-id mixedbread-ai/mxbai-embed-large-v1 \
    --model-id mixedbread-ai/mxbai-rerank-xsmall-v1 \
    --port 7997

• Create a .env file by copying copy from compose.env set up relavant fields. You will need to provide EMBEDDING_SVC_URL and RERANKER_SVC_URL in .env file respectively which will be `http://localhost:7997"

• Now we index the data (sample-data/creditcards) by executing the following command from project root:

  python -m local.ingest
  

• You can also start a FastAPI server: uvicorn --host 0.0.0.0 --port 8000 backend.server.app:app --reload Then, Swagger doc will be available at: http://localhost:8000/ For local version you need not create data sources, collection or index them using API, as it is taken care by local.metadata.yaml and ingest.py file. You can directly try out retrievers endpoint.

• To use frontend UI for quering you can go to : cd frontend and execute yarn dev to start the UI and play around. Refer more at frontend README. You can then query the documents using the UI hosted at http://localhost:5000/

These commands make use of local.metadata.yaml file where you setup qdrant collection name, different data source path, and embedder configurations. You can try out different retrievers and queries by importing them from from backend.modules.query_controllers.example.payload in run.py. To run the query execute the query script from project root: python -m local.run

Overall the architecture of Cognita is composed of several entities

1. Data Sources - These are the places that contain your documents to be indexed. Usually these are S3 buckets, databases, TrueFoundry Artifacts or even local disk

2. Metadata Store - This store contains metadata about the collection themselves. A collection refers to a set of documents from one or more data sources combined. For each collection, the collection metadata stores

   • Name of the collection
   • Name of the associated Vector DB collection
   • Linked Data Sources
   • Parsing Configuration for each data source
   • Embedding Model and Configuration to be used

3. LLM Gateway - This is a central proxy that allows proxying requests to various Embedding and LLM models across many providers with a unified API format. This can be OpenAIChat, OllamaChat, or even TruefoundryChat that uses TF LLM Gateway.

4. Vector DB - This stores the embeddings and metadata for parsed files for the collection. It can be queried to get similar chunks or exact matches based on filters. We are currently supporting Qdrant and SingleStore as our choice of vector database.

5. Indexing Job - This is an asynchronous Job responsible for orchestrating the indexing flow. Indexing can be started manually or run regularly on a cron schedule. It will

   • Scan the Data Sources to get list of documents
   • Check the Vector DB state to filter out unchanged documents
   • Downloads and parses files to create smaller chunks with associated metadata
   • Embeds those chunks using the AI Gateway and puts them into Vector DB

     The source code for this is in the backend/indexer/

6. API Server - This component processes the user query to generate answers with references synchronously. Each application has full control over the retrieval and answer process. Broadly speaking, when a user sends a request

   • The corresponsing Query Controller bootstraps retrievers or multi-step agents according to configuration.
   • User's question is processed and embedded using the AI Gateway.
   • One or more retrievers interact with the Vector DB to fetch relevant chunks and metadata.
   • A final answer is formed by using a LLM via the AI Gateway.
   • Metadata for relevant documents fetched during the process can be optionally enriched. E.g. adding presigned URLs.

     The code for this component is in backend/server/

1. A Cron on some schedule will trigger the Indexing Job
2. The data source associated with the collection are scanned for all data points (files)
3. The job compares the VectorDB state with data source state to figure out newly added files, updated files and deleted files. The new and updated files are downloaded
4. The newly added files and updated files are parsed and chunked into smaller pieces each with their own metadata
5. The chunks are embedded using embedding models like text-ada-002 from openai or mxbai-embed-large-v1 from mixedbread-ai
6. The embedded chunks are put into VectorDB with auto generated and provided metadata

1. Users sends a request with their query

2. It is routed to one of the app's query controller

3. One or more retrievers are constructed on top of the Vector DB

4. Then a Question Answering chain / agent is constructed. It embeds the user query and fetches similar chunks.

5. A single shot Question Answering chain just generates an answer given similar chunks. An agent can do multi step reasoning and use many tools before arriving at an answer. In both cases, the API server uses LLM models (like GPT 3.5, GPT 4, etc)

6. Before returning the answer, the metadata for relevant chunks can be updated with things like presigned urls, surrounding slides, external data source links.

7. The answer and relevant document chunks are returned in response.

   Note: In case of agents the intermediate steps can also be streamed. It is up to the specific app to decide.

Entire codebase lives in backend/

.
|-- Dockerfile
|-- README.md
|-- __init__.py
|-- backend/
|   |-- indexer/
|   |   |-- __init__.py
|   |   |-- indexer.py
|   |   |-- main.py
|   |   `-- types.py
|   |-- modules/
|   |   |-- __init__.py
|   |   |-- dataloaders/
|   |   |   |-- __init__.py
|   |   |   |-- loader.py
|   |   |   |-- localdirloader.py
|   |   |   `-- ...
|   |   |-- embedder/
|   |   |   |-- __init__.py
|   |   |   |-- embedder.py
|   |   |   -- mixbread_embedder.py
|   |   |   `-- embedding.requirements.txt
|   |   |-- metadata_store/
|   |   |   |-- base.py
|   |   |   |-- client.py
|   |   |   `-- truefoundry.py
|   |   |-- parsers/
|   |   |   |-- __init__.py
|   |   |   |-- parser.py
|   |   |   |-- pdfparser_fast.py
|   |   |   `-- ...
|   |   |-- query_controllers/
|   |   |   |-- default/
|   |   |   |   |-- controller.py
|   |   |   |   `-- types.py
|   |   |   |-- query_controller.py
|   |   |-- reranker/
|   |   |   |-- mxbai_reranker.py
|   |   |   |-- reranker.requirements.txt
|   |   |   `-- ...
|   |   `-- vector_db/
|   |       |-- __init__.py
|   |       |-- base.py
|   |       |-- qdrant.py
|   |       `-- ...
|   |-- requirements.txt
|   |-- server/
|   |   |-- __init__.py
|   |   |-- app.py
|   |   |-- decorators.py
|   |   |-- routers/
|   |   `-- services/
|   |-- settings.py
|   |-- types.py
|   `-- utils.py

Cognita goes by the tagline -

Everything is available and Everything is customizable.

Cognita makes it really easy to switch between parsers, loaders, models and retrievers.

• You can write your own data loader by inherting the BaseDataLoader class from backend/modules/dataloaders/loader.py

• Finally, register the loader in backend/modules/dataloaders/__init__.py

• Testing a dataloader on localdir, in root dir, copy the following code as test.py and execute it. We show how to test an existing LocalDirLoader here:

  from backend.modules.dataloaders import LocalDirLoader
  from backend.types import DataSource
  
  data_source = DataSource(
  type="local",
  uri="sample-data/creditcards",
  )
  
  loader = LocalDirLoader()
  
  
  loaded_data_pts = loader.load_full_data(
      data_source=data_source,
      dest_dir="test/creditcards",
  )
  
  
  for data_pt in loaded_data_pts:
      print(data_pt)

• The codebase currently uses OpenAIEmbeddings you can registered as default.
• You can register your custom embeddings in backend/modules/embedder/__init__.py
• You can also add your own embedder an example of which is given under backend/modules/embedder/mixbread_embedder.py. It inherits langchain embedding class.

• You can write your own parser by inherting the BaseParser class from backend/modules/parsers/parser.py

• Finally, register the parser in backend/modules/parsers/__init__.py

• Testing a Parser on a local file, in root dir, copy the following code as test.py and execute it. Here we show how we can test existing MarkdownParser:

  import asyncio
  from backend.modules.parsers import MarkdownParser
  
  parser = MarkdownParser()
  chunks =  asyncio.run(
      parser.get_chunks(
          filepath="sample-data/creditcards/diners-club-black.md",
      )
  )
  print(chunks)

• To add your own interface for a VectorDB you can inhertit BaseVectorDB from backend/modules/vector_db/base.py

• Register the vectordb under backend/modules/vector_db/__init__.py

• Rerankers are used to sort relavant documents such that top k docs can be used as context effectively reducing the context and prompt in general.
• Sample reranker is written under backend/modules/reranker/mxbai_reranker.py

Code responsible for implementing the Query interface of RAG application. The methods defined in these query controllers are added routes to your FastAPI server.

• Add your Query controller class in backend/modules/query_controllers/

• Add query_controller decorator to your class and pass the name of your custom controller as argument

from backend.server.decorator import query_controller

@query_controller("/my-controller")
class MyCustomController():
    ...

• Add methods to this controller as per your needs and use our http decorators like post, get, delete to make your methods an API

from backend.server.decorator import post

@query_controller("/my-controller")
class MyCustomController():
    ...

    @post("/answer")
    def answer(query: str):
        # Write code to express your logic for answer
        # This API will be exposed as POST /my-controller/answer
        ...

• Import your custom controller class at backend/modules/query_controllers/__init__.py

...
from backend.modules.query_controllers.sample_controller.controller import MyCustomController

As an example, we have implemented sample controller in backend/modules/query_controllers/example. Please refer for better understanding

To be able to Query on your own documents, follow the steps below:

1. Register at TrueFoundry, follow here

   • Fill up the form and register as an organization (let's say <org_name>)
   • On Submit, you will be redirected to your dashboard endpoint ie https://<org_name>.truefoundry.cloud
   • Complete your email verification
   • Login to the platform at your dashboard endpoint ie. https://<org_name>.truefoundry.cloud

   Note: Keep your dashboard endpoint handy, we will refer it as "TFY_HOST" and it should have structure like "https://<org_name>.truefoundry.cloud"

2. Setup a cluster, use TrueFoundry managed for quick setup

   • Give a unique name to your Cluster and click on Launch Cluster
   • It will take few minutes to provision a cluster for you
   • On Configure Host Domain section, click Register for the pre-filled IP
   • Next, Add a Docker Registry to push your docker images to.
   • Next, Deploy a Model, you can choose to Skip this step

3. Add a Storage Integration

4. Create a ML Repo

   • Navigate to ML Repo tab

   • Click on + New ML Repo button on top-right

   • Give a unique name to your ML Repo (say 'docs-qa-llm')

   • Select Storage Integration

   • On Submit, your ML Repo will be created

     For more details: link

5. Create a Workspace

   • Navigate to Workspace tab
   • Click on + New Workspace button on top-right
   • Select your Cluster
   • Give a name to your Workspace (say 'docs-qa-llm')
   • Enable ML Repo Access and Add ML Repo Access
   • Select your ML Repo and role as Project Admin
   • On Submit, a new Workspace will be created. You can copy the Workspace FQN by clicking on FQN.

   For more details: link

6. Deploy RAG Application

   • Navigate to Deployments tab
   • Click on + New Deployment buttton on top-right
   • Select Application Catalogue
   • Select your workspace
   • Select RAG Application
   • Fill up the deployment template
     • Give your deployment a Name
     • Add ML Repo
     • You can either add an existing Qdrant DB or create a new one
     • By default, main branch is used for deployment (You will find this option in Show Advance fields). You can change the branch name and git repository if required.

       Make sure to re-select the main branch, as the SHA commit, does not get updated automatically.

     • Click on Submit your application will be deployed.

The following steps will showcase how to use the cognita UI to query documents:

1. Create Data Source

   • Click on Data Sources tab
   • Click + New Datasource
   • Data source type can be either files from local directory, web url, github url or providing Truefoundry artifact FQN.
     • E.g: If Localdir is selected upload files from your machine and click Submit.
   • Created Data sources list will be available in the Data Sources tab.

2. Create Collection

   • Click on Collections tab
   • Click + New Collection
   • Enter Collection Name
   • Select Embedding Model
   • Add earlier created data source and the necessary configuration
   • Click Process to create the collection and index the data.

3. As soon as you create the collection, data ingestion begins, you can view it's status by selecting your collection in collections tab. You can also add additional data sources later on and index them in the collection.

4. Response generation

   • Select the collection
   • Select the LLM and it's configuration
   • Select the document retriever
   • Write the prompt or use the default prompt
   • Ask the query

Your contributions are always welcome! Feel free to contribute ideas, feedback, or create issues and bug reports if you find any! Before contributing, please read the Contribution Guide.

Contributions are welcomed for the following upcoming developments:

• Support for other vector databases like Chroma, Weaviate, etc
• Support for Scalar + Binary Quantization embeddings.
• Support for RAG Evalutaion of different retrievers.
• Support for RAG Visualization.
• Support for conversational chatbot with context
• Support for RAG optimized LLMs like stable-lm-3b, dragon-yi-6b, etc
• Support for GraphDB