Skip to content

01-repl-functions.md

Latest commit

 

History

History
88 lines (54 loc) · 3.08 KB

01-repl-functions.md

File metadata and controls

88 lines (54 loc) · 3.08 KB
Raw

SPEC-01: Advanced REPL Functions

Overview

Add three new REPL helper functions from recurse: map_reduce(), find_relevant(), and extract_functions().

Requirements

map_reduce()

[SPEC-01.01] The system SHALL provide a map_reduce(content, map_prompt, reduce_prompt, n_chunks, model) function.

[SPEC-01.02] map_reduce() SHALL partition content into n_chunks roughly equal parts.

[SPEC-01.03] map_reduce() SHALL apply map_prompt to each chunk in parallel using llm_batch().

[SPEC-01.04] map_reduce() SHALL combine map results and apply reduce_prompt to synthesize final output.

[SPEC-01.05] map_reduce() SHALL support optional model parameter with values: "fast", "balanced", "powerful", "auto" (default).

[SPEC-01.06] map_reduce() SHALL handle content exceeding 1M characters without failure.

find_relevant()

[SPEC-01.07] The system SHALL provide a find_relevant(content, query, top_k, use_llm_scoring) function.

[SPEC-01.08] find_relevant() SHALL partition content into ~50-line chunks with 5-line overlap.

[SPEC-01.09] find_relevant() SHALL perform keyword pre-filtering to identify candidate chunks.

[SPEC-01.10] find_relevant() SHALL optionally use LLM scoring when use_llm_scoring=True and candidates exceed top_k * 2.

[SPEC-01.11] find_relevant() SHALL return a list of (chunk, score) tuples sorted by relevance descending.

[SPEC-01.12] find_relevant() SHALL complete within 2 seconds for content under 100K characters (without LLM scoring).

extract_functions()

[SPEC-01.13] The system SHALL provide an extract_functions(content, language) function.

[SPEC-01.14] extract_functions() SHALL support languages: "python", "go", "javascript", "typescript".

[SPEC-01.15] extract_functions() SHALL return a list of dicts with keys: "name", "signature", "start_line", "end_line".

[SPEC-01.16] extract_functions() SHALL use regex patterns appropriate for each language.

[SPEC-01.17] extract_functions() SHALL handle malformed input gracefully, returning partial results.

Interface

def map_reduce(
    content: str,
    map_prompt: str,
    reduce_prompt: str,
    n_chunks: int = 4,
    model: str = "auto"
) -> str:
    """Apply map-reduce pattern to large content."""

def find_relevant(
    content: str,
    query: str,
    top_k: int = 5,
    use_llm_scoring: bool = False
) -> list[tuple[str, float]]:
    """Find sections most relevant to query."""

def extract_functions(
    content: str,
    language: str = "python"
) -> list[dict]:
    """Extract function definitions from code."""

Security

[SPEC-01.18] All functions SHALL execute within the existing RestrictedPython sandbox.

[SPEC-01.19] map_reduce() and find_relevant() SHALL respect existing budget limits for LLM calls.

Testing Requirements

[SPEC-01.20] Unit tests SHALL cover all functions with edge cases (empty content, single chunk, malformed input).

[SPEC-01.21] Property tests SHALL verify map_reduce() produces consistent results across different chunking strategies.

[SPEC-01.22] Integration tests SHALL verify functions work within REPL environment.