analyxa 0.1.0

Multi-dimensional extraction engine for AI conversations

Analyxa

Multi-dimensional extraction engine for AI conversations.

Analyxa takes opaque conversations between users and AI agents and decomposes them into N configurable dimensions — sentiment, intensity, topics, risk signals, intent, entities, and more — stored as 1,536-dimensional semantic vectors.

---

What it does

Conversation → Analyxa → Structured JSON (N fields) + Semantic Vector (1536D)

One conversation in, structured intelligence out:

• 10 universal fields extracted from any conversation (sentiment, topics, risk signals, intent, entities, action items...)
• Vertical schemas add domain-specific fields: support (16), sales (16), coaching (18)
• Semantic vectors enable similarity search across thousands of conversations
• Pipeline ready: Redis queue → Analyxa → Qdrant vector DB

Quick Start

Installation

pip install analyxa

Python API

from analyxa import analyze

result = analyze(
    "User: I was charged twice for my subscription.\n"
    "Agent: I see the duplicate charge. Processing a refund now.\n"
    "User: Thanks, but please make sure it doesn't happen again.",
    schema="support"
)

print(result.fields["sentiment"])           # "negative"
print(result.fields["satisfaction_prediction"])  # "dissatisfied"
print(result.fields["issue_category"])      # "billing"
print(result.fields["risk_signals"])        # ["frustration", "repeat_contact"]

CLI

# Analyze a conversation file
analyxa analyze conversation.txt --schema support --output result.json

# List available schemas
analyxa schemas list

# Show schema fields
analyxa schemas show support

# Batch analyze a directory
analyxa batch ./conversations/ --schema universal --output-dir ./results/

Environment Setup

Create a .env file:

ANTHROPIC_API_KEY=sk-ant-...      # Required for analysis
OPENAI_API_KEY=sk-...              # Optional, for embeddings
ANALYXA_PROVIDER=anthropic         # or "openai"
ANALYXA_SCHEMA=universal           # Default schema

Schemas

Analyxa uses YAML schemas to define what to extract. Schemas are hierarchical — vertical schemas inherit all universal fields.

Schema	Fields	Description
universal	10	Base fields for any conversation
support	16	Customer support (+satisfaction, issue category, effort score...)
sales	16	Sales conversations (+buying stage, objections, budget signals...)
coaching	18	Coaching/therapeutic (+emotional valence, behavioral patterns, coping strategies...)

Universal Fields (included in all schemas)

Field	Type	Description
title	string	Descriptive session name
summary	string	3-5 sentence summary (vectorized for search)
sentiment	keyword	User sentiment: positive, negative, mixed, neutral
sentiment_intensity	keyword	low, medium, high
topics	keyword_array	Specific topics discussed
session_outcome	keyword	resolved, unresolved, escalated, abandoned
user_intent	string	What the user really needed
risk_signals	keyword_array	frustration, churn_risk, complaint, urgency...
key_entities	keyword_array	People, products, dates, amounts mentioned
action_items	string_array	Explicit commitments or next steps

Custom Schemas

Create your own schema by inheriting from universal:

metadata:
  name: my_vertical
  version: "1.0"
  description: "Custom schema for my use case"
  inherits: universal

fields:
  - name: custom_field
    type: keyword
    required: true
    description: "My custom dimension"
    prompt_hint: "Instructions for the LLM on how to extract this field"
    allowed_values: [option_a, option_b, option_c]

Production Pipeline

Redis → Analyxa → Qdrant

# Start infrastructure
cd docker && docker compose up -d

# Push conversations to Redis queue
analyxa redis push conversation.txt --schema support

# Process all pending conversations
analyxa redis process

# Search by semantic similarity
analyxa search "frustrated customer with billing issue" --limit 5

Python Pipeline

from analyxa.sources.redis_source import RedisSource
from analyxa.sinks.qdrant_sink import QdrantSink
from analyxa.batch import batch_analyze_from_redis

# Process Redis queue → Qdrant
result = batch_analyze_from_redis()
print(f"Processed: {result.successful}/{result.total}")

# Search similar conversations
sink = QdrantSink()
similar = sink.search_similar(query_embedding, limit=10, filters={"sentiment": "negative"})

Configuration

All settings via environment variables or .env file:

Variable	Default	Description
ANTHROPIC_API_KEY	—	Anthropic API key
OPENAI_API_KEY	—	OpenAI API key (for embeddings)
ANALYXA_PROVIDER	anthropic	LLM provider: anthropic or openai
ANALYXA_MODEL	(provider default)	Model override
ANALYXA_SCHEMA	universal	Default schema
ANALYXA_EMBEDDINGS	true	Enable/disable embeddings
REDIS_URL	redis://localhost:6379	Redis connection
QDRANT_URL	http://localhost:6333	Qdrant connection

Architecture

src/analyxa/
├── analyzer.py          # Pipeline orchestrator
├── schema.py            # YAML schema loader with inheritance
├── prompt_builder.py    # Dynamic prompt generation from schemas
├── llm_client.py        # Multi-provider LLM abstraction
├── embeddings.py        # Semantic vector generation (1536D)
├── config.py            # Centralized configuration
├── cli.py               # Click CLI
├── batch.py             # Batch processing
├── sources/
│   ├── file_source.py   # Read from files
│   └── redis_source.py  # Read from Redis queue
├── sinks/
│   ├── json_sink.py     # Write to JSON files
│   ├── stdout_sink.py   # Print to terminal
│   └── qdrant_sink.py   # Store in Qdrant
└── schemas/
    ├── universal.yaml   # 10 base fields
    ├── support.yaml     # +6 support fields
    ├── sales.yaml       # +6 sales fields
    └── coaching.yaml    # +8 coaching fields

License

Apache 2.0 — see LICENSE for details.

Contributing

See CONTRIBUTING.md for guidelines.

---

Built by Next AI Ecosystem