parfold

Parallel async primitives for tree-based operations: fold, unfold, map, filter, and sorting.

Why?

When you need to process many items with async functions (like LLM calls), sequential execution is slow. parfold provides tree-structured parallelism that achieves O(log n) depth instead of O(n).

# Sequential: O(n) round trips
result = items[0]
for item in items[1:]:
    result = await combine(result, item)

# Parallel fold: O(log n) round trips
from parfold import fold
result = await fold(items, combine)

Installation

pip install parfold

Primitives

fold — parallel tree reduction

Combines a list into a single result using parallel binary tree reduction.

from parfold import fold

async def combine(a: str, b: str) -> str:
    # Your async combining logic (e.g., LLM call)
    return await summarize_together(a, b)

chunks = ["chunk1", "chunk2", "chunk3", "chunk4"]
summary = await fold(chunks, combine)

How it works: Instead of combining left-to-right, fold pairs items and combines each pair in parallel, then pairs the results, and so on. For 8 items, that's 3 levels of parallelism instead of 7 sequential operations.

unfold — parallel tree expansion

Expands a seed into leaves by recursively decomposing, with all children expanded in parallel.

from parfold import unfold

async def decompose(query: str) -> list[str]:
    if is_specific_enough(query):
        return []  # Leaf node
    return await generate_subqueries(query)

specific_queries = await unfold("broad research question", decompose)

map — parallel transform

Applies an async function to each item in parallel.

from parfold import map

results = await map(items, async_transform)

filter — parallel predicate

Keeps items where the async predicate returns True.

from parfold import filter

relevant = await filter(items, async_is_relevant)

Sorting

Sorting algorithms using async comparison functions:

from parfold import quicksort, mergesort

async def compare(a, b) -> int:
    # Return negative if a < b, positive if a > b, 0 if equal
    return await llm_compare(a, b)

sorted_items = await quicksort(items, compare)
# or
sorted_items = await mergesort(items, compare)

Quicksort parallelizes comparisons to the pivot within each partition. Mergesort parallelizes the recursive sorting of left/right halves.

BST — Binary Search Tree

Lock-free BST with parallel inserts and O(1) sorted traversal.

from parfold import BST

async def compare(a: str, b: str) -> int:
    return await llm_compare(a, b)

tree = BST(compare)

# Parallel inserts
await asyncio.gather(*[tree.insert(x) for x in items])

# O(1) access to sorted order (no comparisons needed)
for item in tree:
    print(item)

# O(1) min/max
print(tree.min, tree.max)

Uses optimistic concurrency control for parallel inserts. Maintains a threaded linked list through nodes for cheap traversal.

CachedCompare

Wrap your comparison function to cache results:

from parfold import BST, CachedCompare

cached = CachedCompare(llm_compare)
tree = BST(cached)

# After operations:
print(f"Cache: {cached.hits} hits, {cached.misses} misses")

Use Cases

• Summarization: Fold document chunks into a single summary
• Search: Fold chunks while filtering by relevance to a query
• Research expansion: Unfold a broad question into specific searches
• Ranking: Sort items using LLM-based comparison
• Clustering: Fold items into groups using LLM-based merging

Requirements

• Python 3.10+
• No dependencies (just asyncio)

License

MIT