Internal AI R&D · Built by Nocta/case/pia

Pia.
pleiadi.ai

Our in-house research project on cited, sourceable LLM reasoning over structured knowledge. We use Pia to harden the retrieval and evaluation patterns we then apply to client RAG and agent engagements.

Internal AI R&D  Active researchNot a service
Why it exists — § 01

RAG is easy to demo and hard to trust.

Retrieval-augmented generation is the default way to ground an LLM in a specific body of knowledge. The problem isn't building the first version — it's the second mile: answers that cite, retrieval that's evaluated, drift that's detected, and a quality bar that holds when the corpus grows.

Pia is the place we work on that. The patterns that survive there are the ones we ship to clients.

Role
Built & researched by Nocta
Year
Status
  Active R&D
Focus
Sourced reasoning · retrieval · evaluation
Type
Internal R&D · not a service Nocta sells
What it does today — § 02

Reads, plans, cites.

Read

Pia ingests a corpus, chunks it with structure-awareness, and indexes it across hybrid lexical + vector retrieval.

Plan

A planning step decomposes the question, decides which retrieval strategies to use, and budgets tokens before generating.

Cite

Every claim in an answer is tied back to source passages. Unsourced claims are refused, not faked.

Technical approach — § 03

Patterns over models.

Indexing
Structure-aware chunking · hybrid lexical + vector · per-doc metadata
Retrieval
Query rewriting · reranking · per-strategy recall budgets
Reasoning
Plan-then-generate · explicit source binding · refusal on low support
Evaluation
Golden Q/A · grounded-claim checks · citation accuracy scoring
Observability
Per-trace cost & latency · retrieval explainers · regression alerts
Models
Provider-agnostic · swappable per task · cost & quality budgets per call
What we've learned — § 04

Findings, in plain speech.

Retrieval beats reasoning.

For knowledge tasks, the biggest gains come from improving what's retrieved — not from a smarter model on top.

Refusal is a feature.

A system that says “I don't have a source for that” builds more trust than one that confidently confabulates.

Evals before models.

A working evaluation harness is worth more than any single model upgrade — and survives every one of them.

Structure carries.

Respecting the document's own structure during indexing changes retrieval quality more than chunk-size tuning ever does.

What this proves — § 05

RAG that's trustworthy enough to put in front of operators.

Pia gives us a place to keep the boundary between “sounds good” and “is correct” honest. The retrieval, planning, citation and evaluation patterns we work out in Pia are the same ones we deploy in client RAG and agent systems.

Next — § 06

If your team needs answers it can trust — let's design the system.

Pia is research, not a product. The methods are how we build RAG and agents for clients. Start with an audit.