infolang 0.2.0

Official Python SDK for InfoLang semantic memory (il-runtime).

InfoLang Python SDK

Official Python client for InfoLang semantic memory. Wraps the il-runtime REST API (Forge-compatible) with one-line construction, typed errors, automatic retries, and ergonomic agent helpers.

Repository: __REPO__. Package: infolang (PyPI).

Install

While the package is private, install from the repo:

pip install "infolang @ git+ssh://[e-mail]/InfoLang-Inc/__REPO__.git@v0.1.0"

Once published:

pip install infolang

Quickstart

from infolang import InfoLang

il = InfoLang.from_api_key("il_live_...")          # managed cloud
result = il.investigate("How does auth middleware work?")
for chunk in result.chunks:
    print(chunk.score, chunk.text)

Three ways to call, depending on your runtime:

# 1. One-shot
chunks = InfoLang.from_api_key("il_live_...").investigate("query").chunks

# 2. Durable client (connection pooling)
with InfoLang.from_session_file() as il:        # OAuth via ~/.config/infolang/session.json
    il.memorize("a fact worth keeping", source="docs/auth.md")

# 3. Async
import asyncio
from infolang import AsyncInfoLang

async def main():
    async with AsyncInfoLang.from_api_key("il_live_...") as il:
        result = await il.recall("auth middleware", top_k=5)
        print(len(result.chunks))

asyncio.run(main())

Authentication

Mode	Constructor	Target
Managed cloud (API key)	InfoLang.from_api_key("il_live_...")	api.infolang.ai
Managed cloud (OAuth)	InfoLang.from_session_file()	api.infolang.ai
Self-hosted dev	InfoLang.from_dev_key("key:namespace")	127.0.0.1:8766
Enterprise mTLS	InfoLang.from_mtls("client.pem", "client-key.pem")	self-hosted origin

Credentials are also read from the environment: INFOLANG_API_KEY, INFOLANG_DEV_KEY, INFOLANG_BASE_URL, INFOLANG_NAMESPACE.

Core API

Method	Purpose
recall(query, *, namespace, top_k, filters, verbose)	Semantic recall
recall_hybrid(query, *, namespace, top_k, tag_filter, candidate_pool, use_hybrid)	Recall over a candidate pool with tag-inclusion ordering
investigate(query, *, namespace_hint, top_k=5)	Agent-style recall
remember(text, *, source, tags, namespace)	Store a memory
remember_batch(items, *, namespace, source)	Store many memories in one call
memorize(content, *, source, tags, namespace)	Alias of remember
forget(memory_id, *, namespace)	Delete a memory
reset_namespace(namespace)	Bulk clear a namespace (list + forget)
list_banks() / list_recent(*, namespace, n)	Introspection
context_pack(query, *, namespace, max_tokens, repo_root)	One-shot context string
ingest_repo(namespace, *, repo_root, ref)	Index a repository
execute(operations)	Batch ops
health.check()	Liveness/readiness

Bulk ingest + hybrid recall (benchmarks / evals)

il = InfoLang.from_dev_key("devsecret:default")   # self-hosted runtime

il.reset_namespace("eval_run")                      # clean slate
il.remember_batch(
    [
        {"text": "Alice: I moved to Berlin in March 2024.",
         "tags": ["alice", "march", "2024", "session_1"]},
        {"text": "Bob: My flight is on the 12th.",
         "tags": ["bob", "session_2"]},
    ],
    namespace="eval_run",
)

hits = il.recall_hybrid(
    "When did Alice move to Berlin?",
    namespace="eval_run",
    top_k=20,
    tag_filter=["march", "2024"],   # restrict to chunks carrying these tags
    candidate_pool=500,
)
for chunk in hits.chunks:
    print(chunk.score, chunk.tags, chunk.text)

recall_hybrid fetches up to candidate_pool IL-ranked candidates, then (when tag_filter is given) re-orders so chunks sharing a tag come first and truncates to top_k. This biases a wide recall sweep toward tagged evidence — handy for evals/benchmarks — without requiring any server-side hybrid op.

Errors

All failures raise a subclass of InfoLangError: AuthenticationError, RateLimitError (with retry_after), NotFoundError, ValidationError, ServerError, plus InfoLangConnectionError for transport failures. Every API error carries status, body, and request_id.

Resilience

recall/remember and friends retry 429 and 5xx with exponential backoff plus full jitter (configurable via max_retries), and honor Retry-After. Timeouts default to connect 5s / read 30s.

Development

pip install -e ".[dev]"
ruff check .
mypy
pytest

The REST contract is pinned in openapi/ (see openapi/IL_RUNTIME_VERSION). Regenerate the typed models with scripts/codegen.sh after bumping the pin.