Open source · Apache-2.0

Stop feeding your AI chunks.

Coalent caches the understanding your LLM builds — not the raw chunks — reuses it across queries and agents, and refreshes only what a source change touched. Matches chunk-RAG's best accuracy at a fraction of the tokens — plus freshness and compounding reuse RAG can't give.

pip install coalent ·  sits on top of the stack you already run

Same answer, a fraction of the tokens — try it live ↓

Works with the tools you already run

QQdrant
ChChroma
pgpgvector
GrGraphRAG
AiOpenAI
AnAnthropic
LgLangGraph
CwCrewAI
MMCP tools
CfConfluence
JJira
SqSQLite
QQdrant
ChChroma
pgpgvector
GrGraphRAG
AiOpenAI
AnAnthropic
LgLangGraph
CwCrewAI
MMCP tools
CfConfluence
JJira
SqSQLite
See it live

Change a source. Watch it stay correct.

Understanding it can act on — and that never goes stale. Pick a question, then change a source.

Understanding · freshserved · cache hit

Customers can be refunded within 30 days; partial refunds are pro-rated.

Window
30 days
Partial
Pro-rated
Auto-approve
Under $500
Sourceszendesk:refund-guidenotion:billing-policy
Edit a source and watch it stay correct.

↑ live — only the unit that used the changed source rebuilds.

Why Coalent

Understanding, reused, and always fresh.

Real understanding

Coalent synthesizes your sources into a structured, decision-ready briefing — summary, claims, facts — not a pile of fragments.

Reused everywhere

Built once, matched by meaning, reused across every query, session, and agent. Less retrieval, less LLM spend, lower latency.

Always fresh

When a source changes, only the understanding that used it is rebuilt — surgically, the moment it changes. Never stale.

Measured, not asserted

Same answers. Always fresh. A fraction of the tokens.

On the structured-regime benchmark (synthetic templates; real embeddings, a deterministic number check, four answer models), Coalent matches naive dense RAG's accuracy at ~a third of the context tokens — routes to the right source ~every time (route@1 ≈ 1.00), answers multi-hop questions RAG can't, and never goes stale.

Naive dense RAG
Accuracybaseline
Context / read126 tok
Multi-hop0%
Freshnessfresh
re-reads the full context every query
Normal cache (raw)
Accuracy= naive
Context / read126 tok
Multi-hop0%
Freshnessstale
no token savings — and goes stale on change
Coalent
Accuracy= naive
Context / read47 tok
Multi-hop100%
Freshnessfresh
⅓ the tokens · route@1 ≈ 1.00 · always fresh

Real news (MultiHopRAG, pre-registered n=605): matches the best chunk-RAG arm at 0.79x its tokens · 95% null honesty

= naive accuracy (structured)  ·  0% stale  ·  0% fewer tokens (structured)  ·  0% of the tokens matches best chunk-RAG arm (real news)  ·  see the benchmark →

Sits on your stack

Bring any retriever. Keep your model.

Coalent is the layer above retrieval — not a replacement. Plug in your vector DB, GraphRAG, tools, or APIs, and your model of choice. It's one call from any agent framework.

QdrantChromapgvectorGraphRAGOpenAIAnthropicLangGraphCrewAIMCP toolsConfluenceJira
from coalent import SemanticCache, QdrantRetriever, LLMSynthesizer, OpenAIProvider

cache = SemanticCache(
    QdrantRetriever(client=qdrant, collection="docs", embed=embed),
    LLMSynthesizer(OpenAIProvider()),
)

ctx = cache.get("what is our refund policy?")   # decision-ready, fresh

Store understanding, not data.

Caches go stale because they store answers without knowing where they came from. Coalent keeps the lineage — so it always knows exactly what to forget.

The Coalent principle

Community

Build it with us.

Questions, ideas, war stories — come hang out. We're building Coalent in the open, and Discord is the fastest way to reach us.

Give your AI living understanding.

Open source, Apache-2.0. Sits on top of what you already run.