Flex — Knowledge Ecosystem for AI Agents

Knowledge Ecosystem for AI Agents

All your knowledge. One interface.

Your AI doesn't need another retrieval API. It already speaks SQL.

Knowledge Cells

The Ecosystem

One interface, many knowledge bases. Each cell is a SQLite file. Same query language. Same tokens. Same graph intelligence. Different knowledge.

claude_code

Every coding session. Decisions, bugs, architecture choices across all projects.

138K chunks · 3,175 sessions

vector-project-docs

Spec-driven development documentation, architecture, patterns, changelogs.

4.6K chunks · 747 docs

infra-project-docs

Architecture documents, changelogs, agentic memories.

2.4K chunks · 341 docs

Data Model

Architecture

A cell is a SQLite file. The table naming convention is the entire protocol. No manifest, no config, no registry.

THE CELL

_raw_*     immutable     content, embeddings
_edges_*   append-only   relationships
_types_*   immutable     classification
_enrich_*  mutable       graph scores

The prefix tells you the mutability contract.
Table names carry the metadata. Views LEFT JOIN
everything onto _raw_chunks automatically.

Graph scores — centrality, hubs, communities —
are pre-computed columns. Zero query-time cost.

Python 3.12 · SQLite · ONNX · numpy · networkx
No external services. No API keys. Fully local.

Three-Phase Pipeline

Retrieval

SQL pre-filters. Numpy reshapes the embedding landscape. SQL composes the results. Smart data, smart agent, dumb pipe.

SQL pre-filter

Any SQL returning chunk IDs.
Every table with a chunk_id
column is vocabulary.
No registration needed.

numpy landscape

Cosine similarity, then
modulation tokens reshape
scores. Operates on the
full landscape, not top-K.

SQL composition

Full SQL on 500 candidates.
Hub boost, community filter,
JOINs against edge tables.
All composable.

MODULATION TOKENS

Vector search returns the same results every time.
Modulation tokens reshape the landscape per query.

  diverse          spread across subtopics
  recent:N         temporal decay
  unlike:TEXT       subtract a concept
  like:id1,id2     "more like these"
  from:T to:T      trajectory through space

// three tokens, one query
vec_ops('_raw_chunks', 'auth architecture',
  'diverse unlike:OAuth recent:7')

Pipeline

1. SQL narrows the corpus

2. numpy reshapes the landscape

3. SQL composes the results

Smart data, smart agent, dumb pipe. Each stage is independently useful.

Extensibility

Modules

A module is tables. Install by creating them with convention prefixes. Uninstall by dropping them. The system discovers what exists.

SOURCE MODULEScompile a format

A chunk is an atom.
Compile any format — sessions,
docs, code — into the substrate.

  flex init
  flex index claude-code
  flex index docpac ./context

Any format. One cell.

EXTENSION MODULESenrich any cell

A module is a table.
Attach intelligence to any chunk
with CREATE TABLE.

  CREATE TABLE _edges_file_identity (
    chunk_id TEXT, file_uuid TEXT
  )

The view generator discovers it.