compass.validation.content.ParseChunksWithMemory#

class ParseChunksWithMemory(text_chunks, num_to_recall=2)[source]#

Bases: object

Iterate through text chunks while caching prior LLM decisions

This helper stores an in-memory cache of prior validation results so each chunk can optionally reuse outcomes from earlier LLM calls. The design supports revisiting a configurable number of preceding text chunks when newer chunks lack sufficient context.

Parameters:

  • text_chunks (list of str) – List of strings, each of which represent a chunk of text. The order of the strings should be the order of the text chunks. This validator may refer to previous text chunks to answer validation questions.

  • num_to_recall (int, optional) – Number of chunks to check for each validation call. This includes the original chunk! For example, if num_to_recall=2, the validator will first check the chunk at the requested index, and then the previous chunk as well. By default, 2.

Methods

parse_from_ind(ind, key, llm_call_callback)

Validate a chunk by consulting current and prior context
async parse_from_ind(ind, key, llm_call_callback)[source]#

Validate a chunk by consulting current and prior context

Cached verdicts are reused to avoid redundant LLM calls when neighboring chunks have already been assessed. If the cache lacks a verdict, the callback is executed and the result stored.

Parameters:

  • ind (int) – Index of the chunk to inspect. Must be less than the number of available chunks.

  • key (str) – JSON key expected in the LLM response. The same key is used to populate the decision cache.

  • llm_call_callback (callable()) – Awaitable invoked with (key, text_chunk) that returns a boolean indicating whether the chunk satisfies the LLM validation check.

Returns:

bool – True if the selected or recalled chunk satisfies the check, False otherwise.