A deterministic knowledge retrieval engine with symbolic reasoning and evidence-grounded LLM augmentation.
🚧 Work in Progress
Mythrix is an experimental system under active development. Its architecture, domain model, APIs, and user experience are evolving.
Mythrix explores how AI-assisted interpretation can work over large, structured knowledge systems and source corpora without making an LLM responsible for retrieving or inventing the underlying evidence.
The core architectural principle is simple:
The deterministic engine finds the evidence. The LLM helps the user interpret and explore it.
A query against a symbol produces ranked convergence hotspots with verbatim source passages and exact citations. A conversational assistant can then inspect and explore the same evidence through read-only tools, while its tool trace remains visible to the user.
Mythrix is a symbolic knowledge retrieval system that connects a structured symbolic model with a corpus of primary reference sources.
It currently explores symbolic interpretation, modeling relationships between:
- Semiotic systems
- Signs
- Traditions
- Manifestations
- Interpretants
- Source documents and passages
Given a symbolic context, Mythrix derives retrieval signals from the structured model and searches an independent corpus of reference material.
The deterministic query engine then:
- Resolves the symbolic context.
- Derives relevant retrieval signals.
- Searches the source corpus.
- Identifies converging evidence.
- Ranks the resulting regions.
- Returns traceable source passages and citations.
An LLM-powered conversational assistant sits above this engine.
It helps users:
- understand difficult or archaic passages;
- explore retrieved evidence;
- compare passages and concepts;
- ask follow-up questions;
- discover connections across sources;
- formulate new questions;
- request additional queries when exploration leads beyond the current evidence.
The LLM therefore acts as an interpretation and knowledge-discovery assistant, not as the primary retrieval engine.
The central design decision in Mythrix is to deliberately separate retrieval from interpretation.
┌───────────────────────────────────────────────┐
│ GENERATIVE LAYER │
│ │
│ LLM Assistant │
│ │
│ Interpretation │
│ • Explain difficult passages │
│ • Compare concepts and sources │
│ • Answer follow-up questions │
│ • Discover connections │
│ • Maintain conversation │
└───────────────────────▲───────────────────────┘
│
Augmented Evidence
│
┌───────────────────────┴───────────────────────┐
│ EVIDENCE AUGMENTATION LAYER │
│ │
│ Grounded synthesis │
│ │
│ • Summarize evidence regions │
│ • Relate retrieved concepts │
│ • Consolidate multiple passages │
│ • Preserve provenance │
│ │
└───────────────────────▲───────────────────────┘
│
Retrieved Evidence
│
┌───────────────────────┴───────────────────────┐
│ RETRIEVAL LAYER │
│ │
│ Deterministic Query Engine │
│ │
│ • Symbolic resolution │
│ • Retrieval │
│ • Convergence │
│ • Ranking │
│ • Provenance │
│ • Citations │
└───────────────────────┬───────────────────────┘
│
┌─────────┴─────────┐
│ │
▼ ▼
┌──────────────┐ ┌───────────────┐
│ SIGN GRAPH │ │ SOURCE CORPUS │
│ │ │ │
│ Signs │ │ Documents │
│ Traditions │ │ Sections │
│ Interpretants│ │ Passages │
│ Relations │ │ References │
└──────────────┘ └───────────────┘
This boundary is intentional.
The retrieval engine owns the question:
What evidence is relevant?
The LLM helps with the question:
What can we understand or discover from that evidence?
This separation allows the system to use each component where it provides the most value:
| Deterministic engine | LLM assistant |
|---|---|
| Symbolic resolution | Language understanding |
| Retrieval | Explanation |
| Convergence | Contextualization |
| Ranking | Comparison |
| Provenance | Exploration |
| Citations | Knowledge discovery |
The goal is not to eliminate LLMs.
It is to avoid making them responsible for determining what evidence exists.
Retrieval alone is not enough for knowledge-intensive systems.
Once relevant evidence has been identified, Mythrix can augment that evidence with LLM-assisted analysis while preserving the original source boundaries.
The augmentation layer operates over retrieved regions, not over the entire corpus.
It can help:
- summarize multiple evidence regions
- identify relationships between retrieved concepts
- explain difficult passages
- compare different sources
- assist exploration without generating unsupported facts
The LLM does not create new evidence.
It transforms retrieved evidence into a more accessible form for human inquiry.
Mythrix is designed to work with potentially large corpora of reference material.
Many LLM-centered RAG architectures allow the model to influence query formulation, retrieval context, and final generation. This approach can be useful, but it gives the model significant influence over which evidence enters the context.
Mythrix explores a different architecture:
Structured symbolic model
↓
Deterministic retrieval signals
↓
Corpus search
↓
Convergence
↓
Ranking
↓
Evidence surface
↓
LLM interpretation and exploration
The deterministic engine reduces a potentially large corpus to a manageable and explainable evidence surface.
This evidence surface can then be augmented, summarized, and explored by generative models without losing provenance.
The LLM then operates over that surface.
This creates a clear trust boundary:
The LLM can interpret the evidence, but it does not define the evidence.
Mythrix does not treat the source corpus as a flat collection of documents.
The symbolic model provides structured context for retrieval.
Conceptually:
Semiotic System
│
▼
Sign
│
├── Tradition
│
├── Manifestation
│
└── Interpretants
│
▼
Retrieval Signals
│
▼
Source Corpus
The symbolic model actively defines the retrieval space.
A query can produce multiple retrieval signals. For example:
The Sun
│
┌─────────────┼─────────────┐
│ │ │
Laughter Child 100
│ │ │
└─────────────┼─────────────┘
│
▼
Corpus retrieval
│
┌──────────┼──────────┐
▼ ▼ ▼
Passage A Passage B Passage C
│ │ │
└──────────┼──────────┘
▼
Convergence
│
▼
Ranking
│
▼
Evidence
Rather than relying only on the similarity between a single query and a single document chunk, Mythrix can identify regions where multiple symbolic concepts converge.
This convergence contributes to ranking and provides a more explicit explanation of why a region surfaced.
The symbolic graph is therefore not merely metadata around the documents.
The symbolic model actively defines the retrieval space.
The LLM's role begins once evidence is available.
A query may surface passages written in:
- archaic language;
- complex prose;
- unfamiliar terminology;
- culturally specific references;
- dense philosophical or symbolic language.
The user can then ask:
"How do these retrieved passages relate to each other?"
The assistant can help explain the passage, clarify terminology, compare sources, and connect concepts across retrieved evidence.
But the evidence itself remains grounded in the underlying source corpus.
The interaction can then become iterative:
Retrieve
↓
Augment evidence
↓
Interpret
↓
Explore
↓
Ask
↓
Retrieve more
This creates a knowledge-discovery loop in which the LLM acts as a bridge between retrieval and human inquiry.
The assistant can also request a new structured query:
User request
↓
LLM Assistant
↓
Structured query intent
↓
Application command
↓
Deterministic Query Engine
↓
New evidence
The LLM does not execute retrieval itself.
It requests an application action, and the normal query execution path remains responsible for producing the results.
This keeps conversational interaction separate from the deterministic query engine while allowing the two layers to work together.
Mythrix supports two complementary paths.
User
↓
Query UI / API
↓
Deterministic Query Engine
↓
Symbolic Model
↓
Corpus Retrieval
↓
Convergence & Ranking
↓
Evidence
User
↓
Chat
↓
LLM Assistant
│
├───────────────┐
│ │
▼ ▼
Inspect Request
Evidence New Query
│ │
│ ▼
│ Application Command
│ │
│ ▼
│ Deterministic Query
│ │
└───────┬───────┘
▼
Evidence
│
▼
LLM Interpretation
│
▼
User
The deterministic query engine remains the center of evidence retrieval regardless of how the user enters the system.
The system first establishes relevant evidence and then uses the LLM to help interpret and explore it.
The LLM does not determine the underlying evidence or replace deterministic retrieval and ranking.
The symbolic model is an active part of query construction and retrieval behavior.
Results remain connected to their underlying source documents and passages.
The LLM can request application actions, but the application remains responsible for executing them.
The system narrows the corpus before asking the LLM to interpret or explore the resulting evidence.
The LLM is an assistant for interpretation and discovery, not an autonomous authority over the knowledge domain.
Generative capabilities operate only over retrieved and validated evidence surfaces, preserving provenance throughout exploration.
Mythrix is currently built with:
- Python
- LangGraph for conversational agent orchestration
- Kuzu for symbolic graph storage
- Chroma for corpus and vector retrieval
- FastAPI for the application API
- CLI for direct interaction
- Web UI for interactive exploration
- Ollama, running locally, for embedding and generation models — keeping model execution replaceable and independent from the retrieval architecture.
The deterministic query engine is designed to remain usable independently of the conversational agent.
Mythrix is developed using specification-driven development, with system requirements and architectural decisions documented alongside the implementation.
- Setup Guide — Run Mythrix locally and explore the system.
- Architecture — How the system is built: the retrieval pipeline, region augmentation, the agent, and the web viewer, with diagrams and links into the code.
- Corpus Ingestion — Load primary-source documents into the vector-searchable corpus.
- System Specification — Goals, non-goals, requirements, architecture, constraints, and end-to-end flows.
- Architecture Decision Records — The reasoning behind key architectural decisions.
🚧 Work in Progress
Mythrix is an active research and engineering project.
Current areas of development include:
- evolving the symbolic domain model;
- improving retrieval and convergence ranking;
- expanding source provenance;
- refining conversational exploration;
- improving agent/application boundaries;
- evaluating interpretation quality;
- improving the web experience;
- documenting architectural decisions;
- expanding automated tests and evaluation.
The architecture and domain model are expected to evolve as the project develops.
Mythrix explores a broader engineering question:
Can an LLM help people interpret and discover knowledge in large symbolic corpora without becoming responsible for retrieving or inventing the underlying evidence?
The current architecture explores one possible answer:
Structured Symbolic Knowledge
+
Deterministic Retrieval
+
Explicit Evidence
+
LLM-Assisted Interpretation
+
Human Exploration
=
Explainable Knowledge Discovery
Mythrix Engine is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).