hashgrid

The Python client for Hashgrid DNA, a search engine that learns from feedback. You add items and search them; you score the results; DNA learns from the scores, so the next searches rank better.

DNA works on embeddings only and never sees your content. This package embeds text and images on your machine, sends DNA only the vectors, and keeps the content with you, in a local index.

pip install 'hashgrid[local]'

[local] adds the local embedders, which run on your CPU with no account. To embed with OpenAI or Voyage instead, install plain hashgrid and use your own key.

Each grid has its own API key (hg-…). Creating a grid needs no key:

from hashgrid.dna import Grid

grid = Grid.create("mine")
print(grid.key.api_key)   # shown only now: keep it, e.g. as $HASHGRID_API_KEY

Add, search, score

from hashgrid.dna import Grid

grid = Grid()  # Index() and DNA(), from the environment: the key from $HASHGRID_API_KEY

grid.add_many({
    "/cities/lisbon": "Lisbon, a hilly coastal capital",
    "/cities/oslo": "Oslo, between fjords and forest",
})

query, results = grid.ask("a capital by the sea", k=2, prefix="/cities/")
results[0].content                    # "Lisbon, a hilly coastal capital"
grid.score(query, results[0], 1.0)    # a good result
grid.score(query, results[1], -0.5)   # a poor one

• Items have a key, a path unique in the grid. Only items are found by searches, and prefix narrows a search to keys that start with it. Without a key, an item is keyed by a hash of its content, so adding the same content again returns the same item.
• Queries have no key. grid.search(query, k) gives the next k results not shown before.
• Scores run from -1 (wrong) to 1 (perfect). You can score a result, or any pair you already know is good or bad. Only scores train DNA.

Each Result has the item's id, key, content (from the index) and your score. grid.results(query) lists a query's results so far, and grid.get and grid.delete take an item or its id; grid.delete_many deletes many in one request. grid.delete_grid() deletes the whole grid, for good, and its content in the index. AsyncGrid does the same with await.

Images

Content is text or an Image, from a file or from bytes. Images need an embedder that takes them:

from hashgrid.dna import Grid
from hashgrid.index import Image, Index

grid = Grid(Index("fastembed:Qdrant/clip-ViT-B-32"))
grid.add(Image("photos/beach.jpg"), key="/photos/beach")
query, results = grid.ask("a sunny beach", k=5)      # text finds images
query, results = grid.ask(Image("another.jpg"))      # and so do images

The index

Index is the local half of a grid. It embeds content, and keeps it in sqlite by DNA id, so results come back with their content. Index() uses the default embedder and ~/.hashgrid/index.sqlite; $HASHGRID_EMBEDDER and $HASHGRID_INDEX change them, and so do its arguments:

Index("voyage:voyage-multimodal-3")
Index(OpenAI("text-embedding-3-small", dimensions=512), path="my-index.sqlite")
Index(path=":memory:")   # keeps nothing

Use one embedder per grid. Vectors from different models do not compare. The index records each grid's embedder and refuses another, and a grid refuses vectors whose length differs from its own.

Other fastembed, OpenAI and Voyage models work too: fastembed:<model>, openai:<model>, voyage:<model>. DNA takes up to 2048 values, so openai:text-embedding-3-large is shortened to 2048 (openai:<model>@<n> sets the length). With OpenAI or Voyage, your content goes to that provider under your own key, and still never to DNA. Your own embedder is any object with a spec, the modalities it takes ({"text"}, {"text", "image"}) and embed(contents, kind), where kind is "document" or "query".

The API, typed

DNA and AsyncDNA have one method per route of the HTTP API, named after it.

from hashgrid.dna import DNA

dna = DNA("hg-…")
item = dna.create_node([0.1, 0.2, 0.3], key="/items/1")
query = dna.create_node([0.1, 0.2, 0.25])
edge = dna.create_edge(query, prefix="/items/")   # the search; None when nothing is left
dna.put_edge(query, edge, 1.0)                    # the score

Ids are 16-character hex strings. A method also takes the Node or Edge itself. Errors are Unauthorized (401), NotFound (404), Conflict (409, the key is taken) and Invalid (422), all subclasses of DNAError.

MCP server

hashgrid-mcp is a local MCP server, over stdio, for agents that work with text and images. It embeds on your machine, sends DNA only the vectors, and keeps the content in the local index.

{
  "mcpServers": {
    "hashgrid": {
      "command": "uvx",
      "args": ["--from", "hashgrid[local]", "hashgrid-mcp"],
      "env": { "HASHGRID_API_KEY": "hg-…" }
    }
  }
}

With Claude Code:

claude mcp add hashgrid -e HASHGRID_API_KEY=hg-… -- uvx --from 'hashgrid[local]' hashgrid-mcp

Without a key, the agent can call create_grid: the server makes a grid and saves its key in the local index, then uses it from then on.

The tools:

• add_item(text | image, key?): add an item, text or the path of an image file.
• search(query | image, k, prefix): the best results, with their text or image path.
• more_results(query_id, k, prefix): the next ones.
• score(query_id, item_id, score): score a result.
• delete_item(item_id): delete an item.
• create_grid(name): make a grid when the server has no key.
• delete_grid(confirm): delete the grid, for good; confirm is its name.

The first call downloads the local model, about 70 MB. For images, set HASHGRID_EMBEDDER=fastembed:Qdrant/clip-ViT-B-32.

The server records each grid's embedder in the index, and refuses another one for that grid.

The hosted server at https://dna.hashgrid.ai/mcp takes vectors, not content. Use it if your agent makes its own embeddings.

License

MIT