Skip to content

Architecture

AgenticThesis architecture

The editable diagram is available as a standalone HTML source.

In v1.1, collection registers an analysis run only for a needs_review event with a usable disclosure; digest and ignored events remain in Radar without starting analysis. Each run is bound to one disclosure.

Layer Responsibility
Collection Discover configured SEC/IR events and preserve exact artifacts before interpretation
Parsing Produce stable, source-addressable sentence, list-item, table-row, and PDF-page spans
Radar Deterministically decide whether an event needs review, belongs in a digest, or is ignored
Retrieval Combine lexical and vector candidates, rerank conditionally, and pack bounded evidence
Analysis Produce a typed claim-by-claim ThesisDelta from supplied evidence only
Integrity Validate evidence IDs, exact quotes, offsets, claims, and falsifiers
Review Pause for the investor to approve, edit, or reject the proposed change
Commit Apply an approved delta only when its base thesis version is still current
  • Application SQLite is the authoritative record for events, artifacts, Radar, runs, thesis heads, snapshots, and revisions.
  • Checkpoint SQLite holds resumable LangGraph execution state separately from application transactions.
  • Embedded Qdrant stores vectors for local retrieval and can be rebuilt from canonical chunks.
  • Raw artifact bytes and hashes make every accepted quote traceable to the fetched source version.

IR change detection fingerprints normalized visible content so dynamic scripts do not generate duplicate disclosures; the stored artifact hash still identifies the original bytes. Vector indexing, retrieval queries, and analysis have explicit timeouts. Checkpoint recovery uses an explicit deserialization allowlist for workflow types.

AgenticThesis has one application-owned LangGraph with six explicit state transitions. It is not a collection of autonomous agents. Retrieval, validation, commits, and conflict protection remain deterministic; models are restricted to embeddings, conditional reranking, and structured comparison.

This keeps Context, Memory, and Workflow State distinct:

  • Context is the temporary, query-conditioned evidence supplied for one claim.
  • Memory is the approved, immutable thesis snapshot and revision history.
  • Workflow State is the recoverable execution record for an active or reviewed run.

SEC and issuer-owned artifacts may become authoritative evidence. Secondary sources can propose a lead but cannot independently authorize a revision. The model cannot promote a lower-authority source by describing it as official.