GoodMem for LlamaIndex

Use GoodMem as a persistent document and retrieval service in LlamaIndex. GoodMem handles chunking, embeddings and optional reranking; the integration returns native NodeWithScore objects for query engines and agents.

pip install 'goodmem-llamaindex>=0.2.3'

The import namespace is llama_index.tools.goodmem. Set GOODMEM_BASE_URL to your server’s REST root and GOODMEM_API_KEY to its API key.

Store and retrieve Documents

Use an existing GoodMem space configured with an embedder:

import os

from goodmem import Goodmem
from llama_index.core.schema import Document
from llama_index.tools.goodmem import (
    GoodMemDocumentIngestor,
    GoodMemRetriever,
    wait_for_memories,
)

with Goodmem(base_url=os.environ["GOODMEM_BASE_URL"],
             api_key=os.environ["GOODMEM_API_KEY"]) as client:
    space_id = os.environ["GOODMEM_SPACE_ID"]
    ids = GoodMemDocumentIngestor(client=client, space_id=space_id).add_documents([
        Document(text="The returns period is 30 days.",
                 metadata={"source": "https://example.com/returns"})
    ])
    wait_for_memories(client, ids, timeout=120)

    retriever = GoodMemRetriever(client=client, space_ids=[space_id])
    for result in retriever.retrieve("How long do I have to return an item?"):
        print(result.score, result.text, result.metadata.get("source"))

add_documents returns accepted IDs without waiting. Waiting is explicit and checks only those IDs. Ordinary empty searches return immediately.

Connect an agent

Use LlamaIndex’s own tool wrapper. Give each collection a useful name and description:

import os

from llama_index.core.tools import RetrieverTool
from llama_index.tools.goodmem import GoodMemRetriever

space_id = os.environ["GOODMEM_SPACE_ID"]
retriever = GoodMemRetriever(space_ids=[space_id])  # uses environment settings
search = RetrieverTool.from_defaults(
    retriever,
    name="returns_policy",
    description="Search the company's returns and refund policies.",
)
# Pass search to a workflow-based ReActAgent or FunctionAgent.

The model supplies the query; the application configures spaces, filters and reranking. The same retriever works with RetrieverQueryEngine and standard LlamaIndex callbacks.

Pass filters=MetadataFilters(...) for supported scalar comparisons, membership tests and nested conditions. Pass reranker_id=... to rerank on the server without an LLM. Sources, custom metadata and Document metadata exclusions survive storage. Framework scores rank higher as more relevant; raw_score preserves the original value.

Retrieval does not raise on problems the server reports in its results; it returns what the server sent. HTTP errors, such as an unknown space, still raise the SDK's exception. If a configured reranker fails (for example, its ID does not exist), GoodMem reports NOT_FOUND and RERANKING_FAILED and still returns the vector-search hits. The retriever returns those hits as vector scores, negated like any vector score, with score_kind="negative_inner_product", partial=True and the server's statuses on each result, and logs a WARNING. A reported problem with no hits returns an empty list, a UserWarning and a WARNING log line naming the statuses.

Async and administrative tools

aretrieve and aadd_documents use the SDK’s AsyncGoodmem directly. Inject async_client to share its connection pool; caller-owned clients remain open.

GoodMemToolSpec supplies optional space and memory management tools. Its retrieval result includes chunks with the server's raw scores, statuses, a partial flag and score_kind (reranker, or negative_inner_product for vector scores, including when reranking failed). File uploads require an explicitly configured directory. Prefer a scoped retriever tool when an agent only needs search.

Every GoodMem ID you or a model pass in (space, memory, embedder, reranker or LLM) must be a UUID. Anything else raises ValueError before a request is sent, because the SDK puts IDs into URL paths, where a value such as ../spaces/<id> would reach a different resource.

See usage and migration for async examples, supported filters and diagnostics, and the changelog for the changes from 0.1. Run pip install -e '.[dev]', then pytest, ruff check llama_index tests and ruff format --check llama_index tests, as CI does. Live tests are opt-in and clean up their own spaces.