GitHub - kreeben/resin: Language model search engine built on a vector database and an anything key/value store.

5 min read Original article ↗

Resin

Resin is a vector space search engine, a vector database and an anything key/value store. It powers efficient string processing, vector operations, and custom storage primitives designed for speed and simplicity. It can produce large language models out of strings and large anything models out of byte arrays.

Highlights

  • Fast key/value storage with page/column readers and writers
  • Practical text analysis utilities for strings, bags of words/chars, and vectors
  • Commandline tools for building and validating lexicons and comparing strings
  • Clean, dependency light design that is easy to extend

Key/Value Column Semantics

ColumnWriter

  • TryPut(TKey key, ReadOnlySpan value)

    • Inserts the key/value only if the key does not exist in the column-wide snapshot (previous pages included).
    • Returns false when the key already exists; otherwise writes to the current page.
    • Triggers page serialization when the page becomes full.

  • PutOrAppend(TKey key, ReadOnlySpan value)

    • If the key exists anywhere in the column, no new key is stored. Instead, values are linked using a fixed-size node (LinkedAddressNode) written to the value stream.
    • Tail-appending order: the original value remains first, followed by each appended value in insertion order. Address entry for the key points to the list head when linking is active.
    • If the key does not exist in the column snapshot, operates at the page level (insert/append within the current page) and may serialize when full.

ColumnReader

  • Get(TKey key)

    • Returns the value for key. If the key’s address entry points to a linked-list head, returns the concatenated bytes of all linked values.
    • Returns ReadOnlySpan<byte>.Empty when the key does not exist.

  • GetMany(TKey key, out int count)

    • Returns a concatenated ReadOnlySpan<byte> of all values linked for key and outputs the number of items via count.
    • When the key points to a single raw value, returns that value and count = 1. If the key does not exist, returns empty and count = 0.

TKey Restrictions for ColumnWriter/ColumnReader

When working with TKey, please adhere to the following restrictions to ensure proper functionality:

  • TKey must be a value type (struct) and implement both IEquatable<TKey> and IComparable<TKey>.
  • Ordering and equality must be stable across sessions. The column-wide key snapshot uses BinarySearch/sorting, so CompareTo must define a strict total order consistent with Equals.
  • Page-level storage operates on long keys. For primitive numeric keys:
    • double and float are stored via their IEEE bit representations.
    • int and long are stored directly.
    • Other TKey types are hashed via GetHashCode() to a long for page-level operations.
  • Recommendation: Use numeric primitives (double, float, int, long) for deterministic ordering and lookup. If using a custom struct, ensure:
    • Equals and CompareTo are consistent and deterministic.
    • GetHashCode() is stable and evenly distributed; collisions affect page-level operations since non-primitive keys are hashed to long.
  • Keys must be comparable across the entire column; duplicate detection relies on the column snapshot and BinarySearch over sorted keys.

Column model and set operations

  • Each column stores any given TKey at most once in its column-wide snapshot (duplicate keys are prevented by both TryPut and PutOrAppend). This makes columns effectively sets of keys, enabling set operations such as union, intersection, and joins across columns. Linked values (via PutOrAppend) attach additional data to the existing key without introducing duplicates.

Storage artefacts: *.key, *.adr, *.val

  • .key (Key stream)
    • Stores the sorted sequence of TKey representations per page/column. Keys are written in fixed-size slots (sizeof(long) per entry for page-level storage) and serialized in page batches. The column-wide snapshot is built by reading and sorting this stream.
  • .adr (Address stream)
    • Stores Address structs aligned with .key entries. Each Address contains Offset and Length:
      • For raw values: points directly into .val (Offset = start of value, Length = byte length).
      • For linked values: points to a LinkedAddressNode head in .val (Length equals node size). The node chain yields multiple values for a single key.
  • .val (Value stream)
    • Stores the actual value bytes and any LinkedAddressNode headers used for linking. Values are appended at the end of the stream; LinkedAddressNodes are also written into .val to form singly linked lists via absolute offsets.

Immutability of value files

  • The .val stream is treated as append-only:
    • Existing bytes are never modified in place.
    • New values are written at the end, preserving previously written offsets.
    • Linking does not rewrite existing values; instead, LinkedAddressNode headers are appended and previous node’s NextOffset is patched by writing a new node and updating pointers via .adr alignment.
  • Benefits:
    • Stable offsets enable safe caching of addresses and efficient read paths.
    • Appending scales linearly, minimizing fragmentation and avoiding in-place mutations.
    • Historical values remain intact; multi-value chains are expressed via node headers rather than overwriting data.

Usage

  • Use Resin.KeyValue for fast on disk structures and efficient read/write key/value sessions.
  • Use Resin.TextAnalysis for StringAnalyzer, VectorOperations, and similarity tooling.
  • Use Resin.WikipediaCommandLine for commandline tools to build/validate lexicons. See detailed CLI usage and setup in Resin.WikipediaCommandLine/README.md.

Contributing

Contributions are welcome! Please open an issue or pull request with clear motivation, tests when applicable, and concise changes.

License

This project is licensed under the MIT License.

Get help or report issues

Report Issues