3.5 Other Select AI Concepts

Covers additional concepts that support advanced Select AI functionality.

Topics:

• Iterative Refinement
  
• MapReduce
  
• Hallucination in LLM
  
• Live AI Hub
  

3.5.1 Iterative Refinement

Iterative refinement is a process of gradually improving a solution or a model through repeated cycles of adjustments based on feedback or evaluation. It starts with an initial approximation, refines it step by step, and continues until the desired accuracy or outcome is achieved. Each iteration builds on the previous one, incorporating corrections or optimizations to move closer to the goal.

In text summary generation, iterative refinement can be useful for processing large files or documents. The process splits the text into manageable-sized chunks, for example, that fit within an LLM's token limits, generates a summary for one chunk, and then improves the summary by sequentially incorporating the following chunks.

Use cases for iterative refinement:

• Best suited for situations where contextual accuracy and coherence are critical, such as when summarizing complex or highly interconnected texts where each part builds on the previous.
• Ideal for smaller-scale tasks where sequential processing is acceptable.

See Summarization Techniques.

3.5.2 MapReduce

In general, the MapReduce programming model enables processing large-volume data by dividing tasks into two phases: Map and Reduce.

• Map: Processes input data and transforms it into key-value pairs.
• Reduce: Aggregates and summarizes the mapped data based on keys. MapReduce performs parallel processing of large data sets.

In the case of Select AI Summarize, MapReduce partitions text into multiple chunks and processes them in parallel and independently, generating individual summaries for each chunk. These summaries are then combined to form a cohesive overall summary.

Use cases for map reduce:

• Best suited for large-scale, parallel tasks where speed and scalability are priorities, such as summarizing very large data sets or documents.
• Ideal for situations where chunk independence is acceptable, and the summaries can be aggregated later.

See Summarization Techniques.

3.5.3 Hallucination in LLM

Hallucination in the context of Large Language Models refers to a phenomenon where the model generates text that is incorrect, nonsensical, or unrelated to the input prompt. Despite being a result of the model's attempt to generate coherent text, these responses can contain information that is fabricated, misleading, or purely imaginative. Hallucination can occur due to biases in training data, lack of proper context understanding, or limitations in the model's training process.

3.5.4 Live AI Hub

The Live AI Hub architecture allows one database to act as the central metadata repository for both local and remote data sources, that is, Oracle and non-Oracle. Select AI uses this architecture by leveraging the metadata to build an augmented prompt that is sent to the user’s chosen LLM, which then generates a federated SQL query. A key benefit of the Live AI Hub is that it enables data to remain in its original location, eliminating the need for data duplication or complex ETL processes.

It supports federated access to diverse external systems such as, BigQuery, Redshift, multi-cloud, or on-premises databases by securely bridging these sources to Autonomous AI Database.