Introduction#

ByteGenie is like LEGO for building AI agents. It offers an API with tightly integrated data sourcing, processing, and model training functionalities, which enable you to build modular and composable data blocks, model layers, and tools, which can be composed together to build specialised datasets, models, and AI agents.

For example, for an investment research use-case, you can take a few data blocks related to company disclosures and finance textbooks, combine them with a few model layers related to data extraction, data structuring, and financial reasoning, and combine them with a few tools like portfolio calculations, and stock price analytics to build an investment research assistant.

For internal sustainability reporting, you could take a few data blocks related to emission factors, a few model layers related to Excel and PDF document processing, and report generation, and combine them with a few tools related to emission calculations to build and AI assistant for sustainability reporting.

In addition to existing data blocks and model layers that are read to use, users can create their own data blocks by sourcing new data, generate new model layers by training new models, and build new agents combining desired data blocks, model layers, and tools, all from a single consistent API.

ByteGenie agents are essentially configuration files that specify what data objects, model layers, and tools does the agent have access to. Once the agent configuration is set, the agent can be turned on and off, as and when needed, allowing you to incur the costs of agent’s services, only when they are needed.

Supported data sources#

Currently supported data sources include:

  • Company disclosures (annual, sustainability reports, press releases, etc.);

  • Web search (free-form web search using traditional search engines);

  • Events (local events from more than 20 event aggregators);

  • Flights (from more than 5 flight aggregators);

  • Ecommerce product prices (from more than 5 ecommerce websites).

Data processing capabilities#

Currently supported data processing capabilities include:

  • PDF, image, excel, video, and audio files;

  • Short- and long-form document processing;

  • OCR and layout parsing for data extraction;

  • Data structuring for both quantitative and qualitative data in the user-defined structure;

  • Automated data verification for extracted data;

  • Meta-data generation for tabular data;

  • Data vectorisation and semantic search over extracted data for ranking and filtering;

  • Estimation of values for specific indicators, based on extracted data;

  • Data standardisation to homogenize data.

Supported models#

Currently supported foundation models include

  • LLaMA 2 (7b, 13b, 70b);

  • Qwen (7b, 14b);

  • Qwen VL (7b).

Available data blocks#

Currently available data blocks include

  • Full text of company sustainability reports for over 10,000 companies;

  • Full text of company annual reports for over 10,000 companies;

  • Quantitative revenue data block containing quantitative revenue information for over 5,000 companies;

  • Qualitative revenue data block containing quantitative revenue information for over 5,000 companies;

  • Emissions data block containing quantitative emissions data for over 3,000 companies;*

  • Emissions data block containing qualitative emissions data for over 3,000 companies;

  • Emission targets data block containing quantitative emission targets for over 4,000 companies;

  • Emission reduction strategies data block containing emission reduction strategies of over 4,000 companies;

  • Climate risk data block containing climate risk exposures for over 4,000 companies.

Available fine-tuned model layers#

Currently available finetuned model layers include

  • Data structuring model layer;

  • Report generation model layer;

  • Chat model layer;

  • Logical reasoning model layer;

  • Financial ratio calculation and interpretation model layer;

  • Sustainability advisory model layer;

  • Climate risk analysis model layer. These model layers are fine-tuned LoRa weights that can be inserted back into the foundation model to
    build a specialised fine-tuned model.

Supported tools#

Currently supported tools (external APIs) include:

  • Web browsing, including interacting with javascript;

  • Web search, including searching within a website;

  • AWS textract API;

  • Google translate API;

  • Database APIs including Google Storage, AWS S3, BigQuery, Pinecone;

  • Data visualisation tools;

  • Code execution tools (for LLM generated code).

ByteGenie Examples#

The rest of this documentation contains various of using ByteGenie for a variety of use-cases. Examples are categorised by use-cases, including web search, document processing, company research, question-answering on data, model training/fine-tuning, etc.

API Documentation#

The final chapter contains documentation for all the endpoints currently available on the API.