ai.hashgrid由 hashgrid.ai 发布v1.3.0更新于 Oct 31, 2026
Search that learns from feedback, over text and images: embeds locally, sends DNA only vectors.
发布方自己的服务：Hashgrid DNA
Hashgrid DNA 的本地 MCP 服务器：基于反馈学习的文本与图片搜索索引，内容留在本机，只上传向量。
Desktop only，通过 STDIO。 STDIO 服务会启动本地进程，因此需要 SourceWeft 桌面宿主。
参照 仓库 中的启动说明。
The Python client for Hashgrid DNA, a search engine that learns from your grades. You store items, ask a question, get the items that fit it best, and grade them with an LLM, a smaller model or your users. Each grade improves the results for every question.
Each grid has its own API key (hg-…). DNA.signup makes a grid, no key needed, and returns its key:
DNA(api_key) opens the grid, and DNA() opens it from $HASHGRID_API_KEY. A grid made this way holds up
to 1024 items, queries included (dna.info().max_items), and each address can sign up a few times a day.
Need more? Write to [email protected].
42 or tools/42. prefix narrows results and listings to keys that
start with it. Results return keys and metadata, not content: map the keys back to your own data.id. To ask the same question again, reopen it with dna.query(id=query.id), so
its grades add up.An item or a query is text (str), a record (a dict), an image (its bytes, or a Path to its file),
or an embedding from the grid's model (a list of floats). DNA embeds the rest itself.
DNA keeps only what it matches on: text and records up to 8 KB, images as a 512 px WebP. dna.get(key)
reads one item back, as kept.
Methods take a key or the Item / Result itself. AsyncDNA and AsyncQuery have the same methods,
awaited. Errors are subclasses of DNAError: Unauthorized (401), Forbidden (403, the grid is full),
NotFound (404), PreconditionFailed (412), TooLarge (413), UnsupportedType (415), Invalid (422)
and TooManyRequests (429).
For agents, DNA has a hosted MCP server at https://dna.hashgrid.ai/mcp.
MIT
来源：README.md，提交 aeb2def
Add an item to search: `text`, or `image`, the path of an image file (images need an image embedder). `key` is a path unique in the grid (e.g. "/docs/intro"), so searches can narrow by prefix; left out, it is a hash of the content, and adding the same content again changes nothing. Returns the item's id and key.
keytextimageSearch the items with a text `query`, or an `image` file: the best `k` (1 to 50) among those whose key starts with `prefix` ("" for any), best first, each with its id, key and text (or image path). Returns them with the `query_id` to pass to more_results and score.
kimagequeryprefixThe next `k` results of a search (1 to 50), not shown before; empty when none are left.
kprefixquery_id必填Score how well item `item_id` fits search `query_id`, from -1 (wrong) to 1 (perfect): a result of the search, or any item. Only scores train DNA; a new score replaces the last.
score必填item_id必填query_id必填Delete an item, its content in the index and its scores.
item_id必填Create a DNA grid for this server, when it has none (no key needed): its API key is saved in the local index, and the other tools use it from then on.
nameDelete this server's grid, every item and score in it and its content in the local index, for good. `confirm` must be the grid's name. Only when the user asks for it.
confirm必填已满足适用检查项的 90%
该评分衡量此收录条目的完整程度，依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价，也不表示服务器安全或当前可用：可连通性是带日期的检查结果，评分也不包含安装量或用户评价。
请在注册中心条目中声明凭据请求头，或保持端点能够响应探测，以便判定其认证方式。
pip install hashgridfrom hashgrid import DNA
api_key = DNA.signup("mine")print(api_key) # shown only now: store it, e.g. as $HASHGRID_API_KEYfrom hashgrid import DNA
dna = DNA()
dna.put("cities/lisbon", "Lisbon, a hilly coastal capital")dna.put("cities/oslo", "Oslo, between fjords and forest")
query = dna.query("a capital by the sea")results = query.results(limit=2, prefix="cities/")results[0].key # "cities/lisbon"query.grade(results[0], 1.0) # a good resultquery.grade(results[1], -0.5) # a poor onefrom pathlib import Path
dna.put("tools/search", {"name": "search", "description": "Search the web"})dna.put("photos/beach", Path("beach.jpg"))dna.put("raw", '{"a": 1}', type="text/plain") # text that looks like JSON