talkpipe 0.14.2

Python internal and external DSL for writing generative AI analytics

Build and iterate on AI workflows efficiently.

TalkPipe is a Python toolkit for creating, testing, and deploying workflows that combine generative AI with your data and tools. Write pipelines in Python (the Pipe API, chaining steps with |) or in ChatterLang (a concise text DSL) — both use the same building blocks, with LLMs as one tool among many. Pipelines are streaming generators, so large datasets flow through without being loaded into memory, and a pipeline defined once can run from Python, Jupyter, Docker, chatterlang_script, or chatterlang_serve.

Who it's for: engineers and researchers who want scriptable AI pipelines — RAG, batch scoring, web ingestion, agents, and automation — without giving up normal Python when they need it. If your primary model is a large prebuilt agent platform or GUI-first orchestration, another tool may fit better (or layer one on top); TalkPipe stays close to code and scripts.

Typical vertical flow: ingest → chunk/embed → index (for example LanceDB) → retrieve → prompt → answer, then expose the same script with chatterlang_serve if you want an API. See Example 5: RAG pipeline for a full walkthrough.

This README is a high-level overview. Use the documentation map below or the documentation hub for depth.

What Can You Do With TalkPipe?

• Chat with LLMs - Create multi-turn conversations with OpenAI, Ollama, or Anthropic models in just a few lines of code
• Process Documents - Extract text from PDFs, analyze research papers, score content relevance
• Build RAG Pipelines - Create end-to-end Retrieval-Augmented Generation workflows with vector databases
• Analyze Web Content - Download web pages (respecting robots.txt), extract readable text, and summarize
• Build Data Pipelines - Chain together data transformations, filtering, and analysis with Unix-like simplicity
• Deploy Anywhere - Run in Jupyter notebooks, as Docker containers, or as standalone Python applications

How TalkPipe Is Organized

Three layers; use any mix of them in one project:

• Pipe and ChatterLang — Chain sources, segments, and sinks with | in Python, or write ChatterLang scripts for the same concepts (easy to drive from env vars and CI).
• AI and data primitives — LLMs, full-text search, and vector databases behind one style of component.
• Pipelines and applications — Higher-level RAG-style pieces plus CLIs and web apps.

For the full story, see Architecture.

Key Applications

These are the entry points for different usage scenarios, from interactive development to production deployment:

• chatterlang_workbench A browser-based IDE for writing, testing, and running ChatterLang scripts: editor with autocomplete and live error checking, real-time execution, pipeline save/load, next-component suggestions, logging, and documentation lookup.

Your browser does not support the video tag.

• chatterlang_script Runs ChatterLang scripts from files or directly from the command line, for batch processing and automation.

• chatterlang_serve Exposes ChatterLang pipelines as REST APIs or web forms, so workflows can be deployed as web services or user-facing endpoints.

• makevectordatabase & serverag Create vector databases from documents and run RAG web servers in two commands—no scripts required.

• chatterlang_reference_browser & chatterlang_reference_generator Browse installed ChatterLang sources and segments interactively, or generate reference documentation for all of them in HTML and text formats.

• talkpipe_plugins View and manage plugins. TalkPipe's plugin system lets developers register their own sources and segments, so the ecosystem can grow through community contributions and domain-specific extensions.

• Container images Pull release images from GitHub Container Registry (multi-platform on each GitHub release).

Quick Start

Requirements: Python 3.11 or newer. Check your version first with python3 --version — it must report 3.11 or higher before installing.

pip install talkpipe

For LLM support, install the provider(s) you need:

pip install talkpipe[openai]    # OpenAI
pip install talkpipe[ollama]    # Ollama
pip install talkpipe[anthropic] # Anthropic Claude
pip install talkpipe[model2vec] # In-process static embeddings (also in [all])
# Or: pip install talkpipe[all]

Any provider works in any example. The examples in this README mostly show source="ollama", but that is just a per-segment parameter: swap in source="openai" or source="anthropic" (with a matching model) on any LLM segment — the RAG helpers take the same choice as embedding_source/completion_source. Different segments in one pipeline can even use different providers. Installing talkpipe[all] includes all provider integrations, so switching or mixing needs no further installs.

Configure API keys and provider URLs via environment variables (for example OPENAI_API_KEY and ANTHROPIC_API_KEY) or ~/.talkpipe.toml. If TalkPipe runs on a different machine than your Ollama server, set TALKPIPE_OLLAMA_SERVER_URL to that host, e.g. export TALKPIPE_OLLAMA_SERVER_URL="http://<ollama host ip>:11434" (a bare host/IP with no scheme or port, like "myollamahost", also works). Note that the model must already be pulled on that server — run ollama pull llama3.2 there, or OLLAMA_HOST=http://<ollama host ip>:11434 ollama pull llama3.2 from your machine. See Configuration for details and ChatterLang $var substitution.

Hello world (no LLM server required):

Note: source="eliza" is not an LLM provider. It is a local, deterministic adapter for experimenting with TalkPipe/ChatterLang syntax and multi-turn flow when you do not yet have Ollama/OpenAI/Anthropic access.

from talkpipe.chatterlang import compiler

script = '| print | llmPrompt[model="Dr. Eliza", source="eliza", multi_turn=True] | print'
chat = compiler.compile(script).as_function(single_in=True, single_out=True)

chat("Hello, my name is Alice.")
chat("What's my name?")

Tip: from talkpipe.chatterlang import compile is equivalent to the compiler.compile(...) style above — both styles work throughout the docs.

Multi-turn chat (requires a local model server):

Prerequisite: this example uses Ollama, a separate application (not just the talkpipe[ollama] Python package). Install and start it from https://ollama.com/download, then pull the model with ollama pull llama3.2. Cloud users can skip Ollama and substitute source="openai" or source="anthropic" (see the commented variants below).

from talkpipe.chatterlang import compiler

script = '| print | llmPrompt[model="llama3.2", source="ollama", multi_turn=True] | print'
# Using OpenAI:    model="gpt-4o-mini", source="openai"      (set OPENAI_API_KEY)
# Using Anthropic: model="claude-haiku-4-5", source="anthropic" (set ANTHROPIC_API_KEY)
chat = compiler.compile(script).as_function(single_in=True, single_out=True)

response = chat("Hello! My name is Alice.")
response = chat("What's my name?")  # Will remember context

RAG at a glance

Index a list of strings, then ask questions against the store (expand with options in Example 5):

from talkpipe.chatterlang import compiler

docs = ["TalkPipe builds AI pipelines with Python or ChatterLang."]
indexer = compiler.compile(
    '| toDict[field_list="_:text"] | makeVectorDatabase[path="tmp://my_kb", embedding_model="nomic-embed-text", embedding_source="ollama", embedding_field="text", overwrite=True]'
).as_function(single_in=False)
indexer(docs)

rag = compiler.compile(
    '| toDict[field_list="_:text"] | ragToText[path="tmp://my_kb", embedding_model="nomic-embed-text", embedding_source="ollama", completion_model="llama3.2", completion_source="ollama", content_field="text", limit=3] | print'
).as_function(single_in=True)
rag("What is TalkPipe?")

No Ollama server? Swap embedding_source="model2vec" and embedding_model="minishlab/potion-base-8M" for offline embeddings (included in talkpipe[all]). The first run downloads the model from Hugging Face (a few files, ~30 MB); after that it's cached and needs no network — see Precache for offline use to pre-download for air-gapped environments. Note: the ragToText completion step still requires an LLM provider. See the model2vec guide.

Indexing large collections inside a container? Building a vector database over thousands of documents (makevectordatabase, build_rag_database) peaks around 1.5–2 GB of memory. On macOS and Windows, containers run inside the podman machine VM, whose default allocation (often 2 GB) is too tight for that — the ingestion is killed silently (exit code 137; podman inspect shows oom=true). Give the VM more room first:

podman machine stop
podman machine set --memory 4096    # MiB; use 8192 for very large collections
podman machine start

Core Components

1. The Pipe API (Internal DSL)

TalkPipe's Pipe API is a Pythonic way to build data pipelines using the | operator to chain components:

from talkpipe.pipe import io
from talkpipe.llm import chat

# Create a pipeline that prompts for input, gets an LLM response, and prints it
pipeline = io.Prompt() | chat.LLMPrompt(model="llama3.2", source="ollama") | io.Print()
pipeline = pipeline.as_function()
pipeline()  # Run the interactive pipeline

Creating Custom Components

Add new functionality with simple decorators:

from talkpipe.pipe import core, io

@core.segment()
def uppercase(items):
    """Convert each item to uppercase"""
    for item in items:
        yield item.upper()

# Use it in a pipeline
pipeline = io.echo(data="hello,world") | uppercase() | io.Print()
result = pipeline.as_function(single_out=False)()

# Output:
# HELLO
# WORLD
# Returns: ['HELLO', 'WORLD']

2. ChatterLang (External DSL)

ChatterLang provides a Unix-like syntax for building pipelines, perfect for rapid prototyping and experimentation. Run a script from the command line with chatterlang_script --script '<script>', or compile it in Python with compiler.compile('<script>') as shown in the examples below:

INPUT FROM echo[data="1,2,hello,3"] | cast[cast_type="int"] | print

Registering Custom Components for ChatterLang

To make the uppercase segment from section 1 available in ChatterLang, register it with a decorator:

from talkpipe.pipe import core
from talkpipe.chatterlang import registry, compiler

@registry.register_segment("uppercase")
@core.segment()
def uppercase(items):
    """Convert each item to uppercase"""
    for item in items:
        yield item.upper()

# Now use it in ChatterLang scripts
script = 'INPUT FROM echo[data="hello,world"] | uppercase | print'
pipeline = compiler.compile(script).as_function(single_out=False)
result = pipeline()

# Output:
# HELLO
# WORLD
# Returns: ['HELLO', 'WORLD']

The @registry.register_segment() decorator makes your component discoverable by ChatterLang's compiler, allowing you to use it in scripts alongside built-in segments.

Key ChatterLang Features

• Variables: Store intermediate results with @variable_name
• Constants: Define reusable values with CONST name = "value"
• Loops: Repeat operations with LOOP n TIMES { ... }
• Multiple Pipelines: Chain workflows with ; or newlines

Detailed Examples

Example 1: Multi-Agent Debate

Reminder: Examples 1–5 show source="ollama", but any installed provider works in every one of them — swap the source/model parameters as described in the Quick Start.

Problem: Run two LLM personas on one seed topic for several rounds. Result: Printed turns accumulated in @conversation.

from talkpipe.chatterlang import compiler

script = """
CONST economist_prompt = "You are an economist. Reply in one sentence.";
CONST psychologist_prompt = "You are a child psychologist. Reply in one sentence.";

INPUT FROM echo[data="The US should give free puppies to all children."] 
    | @topic 
    | accum[variable=@conversation] 
    | print;

LOOP 3 TIMES {
    INPUT FROM @topic 
        | llmPrompt[system_prompt=economist_prompt, model="llama3.2", source="ollama"] 
        | @topic 
        | accum[variable=@conversation] 
        | print;
    
    INPUT FROM @topic 
        | llmPrompt[system_prompt=psychologist_prompt, model="llama3.2", source="ollama"] 
        | @topic 
        | accum[variable=@conversation] 
        | print;
};

INPUT FROM @conversation
"""

pipeline = compiler.compile(script).as_function()
debate = pipeline()  # Watch the debate unfold!

Example 2: Document Stream Evaluation

Problem: Score a stream of JSONL rows with an LLM against a fixed rubric. Result: A pandas DataFrame with extracted scores per row.

import pandas as pd
from talkpipe.chatterlang import compiler

# Sample document data
documents = [
    '{"title": "Dog", "description": "Dogs are loyal companions..."}',
    '{"title": "Cat", "description": "Cats are independent pets..."}',
    '{"title": "Wolf", "description": "Wolves are wild canines..."}'
]

script = """
CONST scorePrompt = "Rate 1-10 how related to dogs this is:";

| loadsJsonl 
| llmScore[system_prompt=scorePrompt, model="llama3.2", source="ollama", set_as="dog_relevance"] 
| setAs[field_list="dog_relevance.score:relevance_score"] 
| toDataFrame
"""

pipeline = compiler.compile(script).as_function(single_in=False, single_out=True)
df = pipeline(documents)
print(df)
# df now contains relevance scores for each document

Example 3: Web Page Analysis

Problem: Fetch a page, strip boilerplate, summarize with an LLM. Result: Model output to stdout (here, three bullet points).

from talkpipe.chatterlang import compiler

script = """
| downloadURL
| htmlToText
| llmPrompt[
    system_prompt="Summarize this article in 3 bullet points",
    model="llama3.2",
    source="ollama"
  ]
| print
"""

analyzer = compiler.compile(script).as_function(single_in=True)
analyzer("http://example.com/")

Example 4: Content Evaluation Pipeline

Problem: Score each article on two axes, keep only strong matches. Result: Printed dicts for items whose best score exceeds a threshold.

from talkpipe.chatterlang import compiler

# Sample article data
articles = [
    '{"title": "New LLM Model Released", "summary": "AI Company announces new LLM with improved reasoning"}',
    '{"title": "Smart Home IoT Devices", "summary": "Review of latest Arduino-based home automation"}',
    '{"title": "Cat Videos Go Viral", "summary": "Funny cats take over social media again"}',
    '{"title": "RAG Systems in Production", "summary": "How companies deploy retrieval-augmented generation"}',
]

script = """
# Define evaluation prompts
CONST ai_prompt = "Rate 0-10 how relevant this is to AI practitioners. Consider mentions of AI, ML, algorithms, or applications.";
CONST iot_prompt = "Rate 0-10 how relevant this is to IoT researchers. Consider hardware, sensors, or embedded systems.";

# Process articles
| loadsJsonl
| concat[fields="title,summary", set_as="full_text"]

# Score for AI relevance
| llmScore[system_prompt=ai_prompt, field="full_text", set_as="ai_eval", model="llama3.2", source="ollama"]
| setAs[field_list="ai_eval.score:ai_score,ai_eval.explanation:ai_reason"]

# Score for IoT relevance
| llmScore[system_prompt=iot_prompt, field="full_text", set_as="iot_eval", model="llama3.2", source="ollama"]
| setAs[field_list="iot_eval.score:iot_score,iot_eval.explanation:iot_reason"]

# Find highest score
| lambda[expression="max(item['ai_score'],item['iot_score'])", set_as="max_score"]

# Filter articles with score > 6
| gt[field="max_score", n=6]

# Format output
| toDict[field_list="title,ai_score,iot_score,max_score"]
| print
"""

evaluator = compiler.compile(script).as_function(single_in=False, single_out=False)
results = evaluator(articles)

# Output shows only relevant articles with their scores. Scores are assigned by the
# LLM, so they (and which rows survive the >6 filter) vary by model and run; with
# some models only one or two of the four articles may pass. Illustrative example:
# {'title': 'New LLM Model Released', 'ai_score': 9, 'iot_score': 2, 'max_score': 9}
# {'title': 'Smart Home IoT Devices', 'ai_score': 3, 'iot_score': 9, 'max_score': 9}
# {'title': 'RAG Systems in Production', 'ai_score': 8, 'iot_score': 2, 'max_score': 8}

Example 5: RAG Pipeline with Vector Database

Problem: Embed texts into a local vector store, then answer questions with retrieval + completion. Result: String answers from ragToText, plus patterns for yes/no (ragToBinaryAnswer) and numeric scores (ragToScore).

from talkpipe.chatterlang import compiler

# Sample knowledge base documents
documents = [
    "TalkPipe is a Python toolkit for building AI workflows. It provides a Unix-like pipeline syntax for chaining data transformations and LLM operations.",
    "TalkPipe supports multiple LLM providers including OpenAI, Ollama, and Anthropic. You can switch between providers easily using configuration.",
    "With TalkPipe, you can build RAG systems, multi-agent debates, and document processing pipelines. It uses Python generators for memory-efficient streaming.",
    "TalkPipe offers two APIs: the Pipe API (internal DSL) for Python code and ChatterLang (external DSL) for concise script-based workflows.",
    "Deployment is flexible with TalkPipe - run in Jupyter notebooks, Docker containers, or as standalone applications. The chatterlang_serve tool creates web APIs from scripts."
]

# First, index your documents into a vector database
indexing_script = """
| toDict[field_list="_:text"]
| makeVectorDatabase[
    path="./my_knowledge_base",
    embedding_model="nomic-embed-text",
    embedding_source="ollama",
    embedding_field="text",
    overwrite=True
  ]
"""
indexer = compiler.compile(indexing_script).as_function(single_in=False)
indexer(documents)

# Now query the knowledge base with RAG
query_script = """
| toDict[field_list="_:text"]
| ragToText[
    path="./my_knowledge_base",
    embedding_model="nomic-embed-text",
    embedding_source="ollama",
    completion_model="llama3.2",
    completion_source="ollama",
    content_field="text",
    prompt_directive="Answer the question based on the background information provided.",
    limit=3
  ]
| print
"""

rag_pipeline = compiler.compile(query_script).as_function(single_in=True)
answer = rag_pipeline("What are the key benefits of using TalkPipe?")
# Returns an LLM-generated answer based on relevant document chunks

# For yes/no questions, use ragToBinaryAnswer:
binary_rag_script = """
| toDict[field_list="_:text"]
| ragToBinaryAnswer[
    path="./my_knowledge_base",
    embedding_model="nomic-embed-text",
    embedding_source="ollama",
    completion_model="llama3.2",
    completion_source="ollama",
    content_field="text"
]
| print
"""
binary_rag = compiler.compile(binary_rag_script).as_function(single_in=True)
result = binary_rag("Does TalkPipe support Docker?")
result = binary_rag("Does TalkPipe have a podcast about pipes?")

# For scored evaluations, use ragToScore:
score_rag_script = """
| toDict[field_list="_:text"]
| ragToScore[
    path="./my_knowledge_base",
    embedding_model="nomic-embed-text",
    embedding_source="ollama",
    completion_model="llama3.2",
    completion_source="ollama",
    prompt_directive="Answer the provided question on a scale of 1 to 5.",
    content_field="text"
  ]
| print
"""
score_rag = compiler.compile(score_rag_script).as_function(single_in=True)
score = score_rag("How flexible is talkpipe?")
score_rag("How well does this text describe pipe smoking?")

Documentation

For comprehensive documentation and examples, see the docs/ directory.

Goal	Start here
Install and first pipeline	Getting started
Commands and components	API reference
Walkthroughs	Tutorials
Design and extending TalkPipe	Architecture
Official container images (GHCR, multi-arch releases)	Container images
Contributor glossary and conventions	Developer handbook

Architecture & Development

Design Principles

• Dual-language architecture — Pure Python (Pipe API) for maximum flexibility and IDE support; ChatterLang for concise, portable scripts.
• Streaming architecture — Python generators throughout: memory-efficient processing of large datasets, real-time results as data flows, natural integration with streaming sources.
• Extensibility first — Simple decorators (@source, @segment, @field_segment) for adding functionality; components are just Python functions, easy to test and debug; mix TalkPipe with any Python code or library.

Project Structure

talkpipe/
├── app/          # Runnable applications (servers, CLIs)
├── chatterlang/  # ChatterLang parser, compiler, and components
├── data/         # Data manipulation and I/O components
├── llm/          # LLM integrations (OpenAI, Ollama, Anthropic)
├── operations/   # Algorithms and data processing
├── pipe/         # Core pipeline infrastructure
├── pipelines/    # High-level pipeline components (RAG, vector DB)
├── search/       # Search engine integrations (Whoosh, LanceDB)
└── util/         # Utility functions and configuration

Configuration

TalkPipe uses a flexible configuration system via ~/.talkpipe.toml or environment variables. For LLM and embedding model / source defaults, see Model and source configuration.

# ~/.talkpipe.toml
default_model_name = "llama3.2"
default_model_source = "ollama"
smtp_server = "smtp.gmail.com"
smtp_port = 587

Environment variables use the TALKPIPE_ prefix:

export TALKPIPE_email_password="your-password"
export OPENAI_API_KEY="sk-..."

Performance Optimization

TalkPipe includes an optional lazy loading feature that can dramatically improve startup performance (up to 18x faster) by deferring module imports until needed:

# Enable lazy loading for faster startup
export TALKPIPE_LAZY_IMPORT=true

This is especially useful for CLI tools and scripts that don't use all TalkPipe features. See the lazy loading documentation for details.

Development Guidelines

Contributor-focused glossary, naming conventions, parameter semantics, standard config keys, and segment/source reference notes live in the developer handbook.

Naming Conventions

• Classes: CamelCase (e.g., LLMPrompt)
• Decorated functions: camelCase (e.g., @segment def extractText)
• ChatterLang names: camelCase (e.g., llmPrompt, toDataFrame)

Creating Components

Sources generate data (see Creating Custom Components above for segments, which transform data):

from talkpipe.pipe import core, io

@core.source()
def fibonacci(n=10):
    a, b = 0, 1
    for _ in range(n):
        yield a
        a, b = b, a + b

# Use it in a pipeline
pipeline = fibonacci(n=5) | io.Print()
result = pipeline.as_function(single_out=False)()

# Output:
# 0
# 1
# 1
# 2
# 3
# Returns: [0, 1, 1, 2, 3]

Field Segments provide a convenient way to create 1:1 segments:

from datetime import datetime
from talkpipe.pipe import core, io
from talkpipe.chatterlang import registry

@registry.register_segment("addTimestamp")
@core.field_segment()
def addTimestamp(item):
    # Handle a single item, not an iterable
    # The decorator handles set_as and field parameters automatically
    return datetime.now()

# Use it with dictionaries
data = [{'name': 'Alice'}, {'name': 'Bob'}]
pipeline = addTimestamp(set_as="timestamp") | io.Print()
    
result = pipeline.as_function(single_in=False, single_out=False)(data)

# Output (timestamps will vary):
# {'name': 'Alice', 'timestamp': datetime.datetime(2024, 1, 15, 10, 30, 45, 123456)}
# {'name': 'Bob', 'timestamp': datetime.datetime(2024, 1, 15, 10, 30, 45, 234567)}

# Now it's also available in ChatterLang:
# script = '| addTimestamp[set_as="timestamp"] | print'

Best Practices

1. Units with side effects should pass data through - e.g., writeFile should yield items after writing
2. Use descriptive parameter names with underscores (e.g., fail_on_error, set_as)
3. Handle errors gracefully - use fail_on_error parameter pattern
4. Document with docstrings - they appear in generated documentation
5. Test with both APIs - ensure components work in both Python and ChatterLang

Roadmap & Contributing

TalkPipe is under active development. Current priorities:

• Enhanced LLM Support: Additional providers, expanded guided generation
• Data Connectors: More database integrations, API clients, file formats
• Workflow Features: Conditional branching, enhanced error handling, retry logic
• Performance: Parallel processing optimization, enhanced lazy loading, better caching
• Developer Tools: Better debugging, testing utilities, IDE plugins
• RAG & Search: Advanced retrieval strategies, hybrid search, multi-modal embeddings

We welcome contributions! Whether it's new components, bug fixes, documentation, or examples, please check our GitHub repository for contribution guidelines.

Status

TalkPipe is in active development: feature-rich and in use, but APIs may evolve. We follow semantic versioning: minor versions aim for compatibility within a major series; major bumps may include breaking changes. Reasonably stable for everyday use: install from PyPI, the | pipeline model, compiler.compile(...).as_function(...), and optional extras for LLM providers.

License

TalkPipe is licensed under the Apache License 2.0. See LICENSE file for details.

---

Last reviewed: 2026-07-24