Inherent
One brain, every agent. Store company context once and let every AI agent
query it with citations, instead of paying to re-load the same context into
every prompt.
Inherent is the open-source context layer behind inherent.sh.
You connect sources (text, Markdown, PDF, DOCX, source code, and
more). It extracts, chunks, embeds, and indexes
them, then serves cited retrieval over REST and MCP. Self-host it in your own
VPC or run it locally with one command.
https://github.com/user-attachments/assets/c01c0e25-c995-4e70-b07c-ce6da2ff393d
Why Inherent
- Connected once, queried everywhere. Ingest a source one time; every agent reads from the same governed index.
- Every answer cites its source. Retrieval returns passages with document, chunk, and position, not a black-box summary.
- Measured, not claimed. Retrieval quality is a CI gate with a ratcheting baseline (table below).
- Yours to run. FastAPI, PostgreSQL, Weaviate, Temporal, Valkey, S3. MIT licensed, deployable in your VPC.
Quickstart
Prerequisites: Docker, Python 3.11+.
pip install inherent
inherent up
inherent connect claude --print
inherent up pulls the version-matched engine images, waits for health, and
stores local credentials in ~/.inherent/.
Contributing from a checkout instead?
git clone https://github.com/inherent-prime/inherent.git && cd inherent
make quickstart # .env, install, Compose stack, dev workspace + API key, health check
make quickstart already seeds a dev workspace and API key (via bootstrap).
Then follow Getting Started Locally for
endpoints, a smoke test, and troubleshooting. Going live? Read
Taking Inherent to Production or
Deploy to Azure.
First search
export API_BASE=http://localhost:18000 API_KEY=ink_dev_local_key_001 WORKSPACE_ID=ws_local_001
# Upload
curl -s -X POST "$API_BASE/v1/documents" -H "X-API-Key: $API_KEY" -H "X-Workspace-Id: $WORKSPACE_ID" \
-F "file=@docs/examples/sample-documents/sample.txt;type=text/plain" | jq .document_id
# Search (once status is "processed")
curl -s -X POST "$API_BASE/v1/search" -H "X-API-Key: $API_KEY" -H "X-Workspace-Id: $WORKSPACE_ID" \
-H "Content-Type: application/json" -d '{"query":"what retrieval modes does Inherent support","limit":3}' | jq .
Full endpoint reference: REST API · MCP tools · CLI · Examples.
Architecture
documents
|
v
+------------------------+
| inh-ingestion-svc |
| extract / chunk / |
| embed / index |
+-----------+------------+
|
+-------------+-------------+
v v
+-------------+ +-------------+
| PostgreSQL | | Weaviate |
| metadata | | vectors |
+------+------+ +------+------+
\ /
v v
+-------------------+
| inh-public-api-svc|
| REST + MCP |
+---------+---------+
|
v
agents
Ingestion consumes upload events from Valkey, runs Temporal workflows, and
writes to PostgreSQL and Weaviate. The public API reads both and serves search
over REST and MCP. Product boundary and non-goals:
ADR 0001.
Retrieval quality baseline
Committed floor in retrieval_baseline.json,
measured on the golden corpus over the full Compose stack. A regression over
0.02 fails the build; a green run on main ratchets the floor up.
| Mode | Recall@5 | MRR | nDCG@5 |
|---|
| Hybrid | 0.910 | 0.885 | 0.844 |
| Keyword | 0.860 | 0.841 | 0.796 |
| Semantic | 0.880 | 0.798 | 0.768 |
Details and local runs: docs/testing.md.
Project health
- Every PR runs lint, format, mypy, bandit, unit tests, Compose E2E, and the retrieval eval gate.
- Coverage floors are enforced and never lowered: ingestion 82%, public API 84%, CLI 89%, contracts 98%, plus per-module floors on auth, search, and verify.
- Releases follow Keep a Changelog and semver; images publish to
ghcr.io/inherent-prime/* on every tag.
- Docs build in strict mode on every PR and publish to docs.inherent.sh.
Repository layout
Contributing and support
License
MIT
