How MyJourney AI works

Every step of the pipeline is designed to be visible, explainable, and demo-friendly.

   ┌───────────┐   ┌──────────────┐   ┌────────────────┐   ┌───────────────┐   ┌────────────────────┐   ┌────────────────────┐
   │  1 Ingest │──▶│ 2 Extract/OCR│──▶│ 3 Classify     │──▶│ 4 Embed (3072)│──▶│ 5 Relationship LLM │──▶│ 6 RAG retrieval    │
   │ PDF/DOCX/ │   │ gemini-2.5   │   │ zero-shot LLM  │   │ gemini-       │   │ cosine + LLM       │   │ hybrid + re-rank + │
   │ image/URL │   │ flash + vis. │   │ + confidence   │   │ embedding-2   │   │ validate edges     │   │ inline citations   │
   └───────────┘   └──────────────┘   └────────────────┘   └───────────────┘   └────────────────────┘   └────────────────────┘
        │                                        │                  │                    │                       │
        ▼                                        ▼                  ▼                    ▼                       ▼
   documents                              entities (with     pgvector HNSW     relationships (typed +      Growth narrative
   (Postgres +                            attributes,        halfvec index     confidence + reasoning)     · Gap analysis
   Storage)                               confidence,        on 3072-dim                                    · Ask your journey
                                          reasoning)         cosine

1. Stage

Ingestion

PDF, DOCX, image, URL, or pasted text. Files land in Lovable Cloud storage; text is captured verbatim.

2. Stage

Extraction & OCR

gemini-2.5-flash reads the document. Scanned images are OCR'd by the same vision-capable model.

3. Stage

Zero-shot classification

Each atomic entity is classified into one of skill · project · certification · internship · achievement · academic, with a 0-1 confidence and one-sentence reasoning.

4. Stage

Embeddings

gemini-embedding-2 → 3072-dim vectors stored in Postgres pgvector, indexed with HNSW halfvec cosine.

5. Stage

Relationship engine

New entities are compared to existing ones by cosine similarity; the LLM validates each candidate edge with a one-sentence reasoning and confidence.

6. Stage

RAG retrieval

Queries are embedded, top-k matched via pgvector, and re-ranked/grounded by the LLM, which cites the retrieved entities and source documents.

Data model

Document (id, type, upload_date, raw_url, extracted_text, pipeline_trace)
Entity (id, type, name, attributes, confidence, reasoning, event_date, source_document_id, embedding)
Relationship (id, source_entity_id, target_entity_id, relation_type, confidence, reasoning)
Embedding pgvector(3072) with HNSW halfvec cosine index