Rel v2

[ngupta10]

Relationship Extraction 1st iteration implementation.

1. The rel extraction engine currently takes 2 entities and the context between them to output a semantic triple (Entity1, Predicate, Entity2), along with {Predicate_type}.
2. Using LLAMA based open source LLMs to extract predicate and predicate_type
3. Using Grammar based approach for ensuring consistent schema of llm output
4. Implemented Branch and Solve Algorithm (currently not using the full implementation i.e. with different llms) - This can be later turned on to extract predicate from different LLMS at same time and validate the llm output.
5. Implement RAG based appraoch for context filtering before feeding it to llm to derive the predicate and predicate_type (turned off for now). Will need to experiment more to determine what are the parameters that we can hypertune to make this more effective.

To-Do:

1. Implement a separate signal for context embeddings - https://github.com/Querent-ai/querent/issues/194
2. bug fixing in entity filtering mechanism of bert llm - https://github.com/Querent-ai/querent/issues/193
3. Establish Fixed & Sample entities and relationships - https://github.com/Querent-ai/querent/issues/195

[ngupta10]

OK sir

[saraswatpuneet]

Still needs some love, let's walk through together