Examples

Working pipelines you can clone and run. Each one reads a source, transforms it in a Python flow, and keeps a target in sync. Start from the closest one and change the parts that differ.

32 examples

Multimodal search5

Search images, PDFs, slides, and faces by meaning: the same vector index with a different encoder.

  1. Local FS
  2. CLIP
  3. Qdrant
  1. Images
  2. ColPali
  3. Qdrant

Image search with ColPali

Embed images and queries into multi-vector ColPali bags of patch vectors, store them in Qdrant, and rank with late-interaction MaxSim through a FastAPI app.

~12 minIntermediate

  1. Local FS
  2. Face detection
  3. Qdrant

Build your own face search

Detect every face in a folder of photos, embed each into a 128-d vector with face_recognition (dlib), and index them in Qdrant — then search your photos by face.

~15 minIntermediate

  1. Local FS
  2. ColPali
  3. Qdrant

Multi-format visual search

Index PDFs and images as page screenshots with the ColPali multi-vector model — no OCR, no chunking — into Qdrant with MaxSim, and search the visual content by text.

~20 minAdvanced

  1. Local FS
  2. Vision LLM + TTS
  3. LanceDB

Slides to narrated search

Render each slide, write speaker notes with a vision LLM, narrate them with Pocket TTS locally on the CPU, and embed the notes into LanceDB — search a deck by meaning and play back the narration.

~25 minAdvanced

Structured extraction6

Extract typed, structured data from code and documents with an LLM, through instructor, BAML, or DSPy.

  1. Multi-repo
  2. Structured output
  3. Local FS

Multi-codebase summarization

Walk many Python repos, LLM-extract a typed summary per file — classes, functions, Mermaid call graphs — and aggregate into an always-fresh wiki page per project.

~35 minAdvanced

  1. HackerNews API
  2. LiteLLM
  3. Postgres
  1. PDF
  2. OpenAI
  3. Postgres

Index academic papers

LLM-extract title, authors, and abstract from academic PDFs into typed rows, embed them, and store it all in Postgres with pgvector.

~12 minIntermediate

  1. PDF
  2. BAML
  3. JSON files

Patient intake forms to typed JSON with BAML

Extract schema-validated patient records from intake-form PDFs with one type-safe BAML Gemini-vision call per file, writing a JSON file per form.

~10 minIntermediate

  1. PDF
  2. Structured extraction
  3. JSON files

Patient intake forms to typed JSON with DSPy

Render patient intake PDFs to images and extract a typed Patient with a DSPy ChainOfThought vision module on Gemini, writing one validated JSON file per form.

~10 minIntermediate

  1. Local FS
  2. docling + LLM
  3. Postgres

Manuals to structured data

Convert PDF manuals to Markdown with docling, LLM-extract a typed module summary (classes, methods, arguments), and store one structured record per manual in Postgres.

~20 minIntermediate

Knowledge graphs4

Give agents a persistent, graph-shaped memory built from conversations, meetings, and products.

  1. YouTube
  2. Entity resolution
  3. SurrealDB

Podcasts to a knowledge graph

Turn YouTube podcasts into a queryable knowledge graph — diarized transcription, two-step LLM extraction, embedding-based entity resolution, and a SurrealDB graph.

~40 minAdvanced

  1. Local FS
  2. Any LLM
  3. Neo4j

Docs to a knowledge graph

Turn a folder of Markdown docs into a Neo4j concept graph — LLM-extracted (subject, predicate, object) triples that stay in sync as the docs change.

~20 minIntermediate

  1. Google Drive
  2. Entity resolution
  3. Neo4j

Meeting notes to a knowledge graph

Turn Google Drive meeting notes into a Neo4j knowledge graph — LLM extraction of organizers, attendees, and tasks, plus embedding-based person entity resolution, kept in sync as notes change.

~25 minIntermediate

  1. Local FS
  2. LLM extraction
  3. Neo4j

Product recommendation graph

LLM-extract what each product is and what pairs with it from product docs, into a Neo4j graph of products and taxonomies that powers "customers also bought" recommendations.

~20 minIntermediate

Ingest and transform6

Move and reshape data with the same declarative flow: your own source, target, or parser.

  1. PDF
  2. Custom blocks
  3. Local FS

PDF to Markdown

Convert a folder of PDFs to Markdown incrementally, with custom building blocks: only new or changed PDFs are converted again.

~18 min

  1. Local FS
  2. Kafka

CSV to Kafka

Watch a folder of CSV files and publish each row as a JSON message to a Kafka topic — declarative target states, only-changed-rows produces, and live mode in ~60 lines.

~12 minStarter

  1. Audio files
  2. Transcription
  3. Postgres

Audio to text

Transcribe local audio files with a LiteLLM speech-to-text model and store one transcript row per file in Postgres, keyed by filename.

~12 minStarter

  1. Postgres
  2. Embeddings

Postgres as a source

Read rows from an existing Postgres table, derive fields, embed each row, and write the vectors back to Postgres with pgvector.

~8 minStarter

  1. Local FS
  2. markdown-it-py

Transform a folder of files

The smallest end-to-end CocoIndex pipeline — watch a folder of Markdown, render each file to HTML, and write the outputs to a local folder incrementally.

~6 minStarter

  1. Kafka
  2. LanceDB

Consume Kafka into LanceDB

Consume JSON messages off a Kafka topic and dispatch each one — by its shape — into the matching LanceDB table.

~10 minIntermediate

Can't find the shape you need?

Clone the closest example, swap the source or the target, and keep the rest. Or open an issue and describe the pipeline.

Read the quickstartRequest an exampleAsk on Discord