chat-search

Chat with documents, search via natural language.

chat-search supports hybrid language models to add chat capabilities to website. RAG built with LangChain, Redis, various model providers (OpenAI, Ollama, vLLM, Huggingface).

Demo: Chat about my blog

Usage

Setup .env

cp .env.example .env

Populate .env file with the required environment variables.

Name	Value	Default
AUTH_TOKEN	auto token used for ingest
CHAT_PROVIDER	model provider, openai or ollama	openai
DEBUG	enable DEBUG, 1 or 0	0
DIGEST_PREFIX	prefix for digest in Redis	digest
DOCUMENT_CONTENT_DESCRIPTION	document content description	Document content
EMBEDDING_DIM	embedding dimensions	1536
EMBEDDING_PROVIDER	embedding provider, openai or ollama or huggingface	openai
ENABLE_FEEDBACK_ENDPOINT	enable feedback endpoint, 1 or 0	1
ENABLE_PUBLIC_TRACE_LINK_ENDPOINT	enable public trace link endpoint, 1 or 0	1
FULLTEXT_RETRIEVER_SEARCH_K	fulltext retriever search result number	4
FULLTEXT_RETRIEVER_SELF_QUERY	whether to enable fulltext retriever self query, 1 or 0	1
FULLTEXT_RETRIEVER_WEIGHT	fulltext retriever weight	0.5
HEADERS_TO_SPLIT_ON	html headers to split text	h1,h2,h3
HF_HUB_EMBEDDING_MODEL	huggingface hub embedding model or Text Embeddings Inference url	http://localhost:8080
INDEX_NAME	index name	document
INDEX_SCHEMA_PATH	index schema path	(will use app/schema.yaml)
LANGCHAIN_API_KEY	langchain api key for langsmith
LANGCHAIN_ENDPOINT	langchain endpoint for langsmith	https://api.smith.langchain.com
LANGCHAIN_PROJECT	langchain project for langsmith	default
LANGCHAIN_TRACING_V2	enable langchain tracing v2	true
LLM_TEMPERATURE	temperature for LLM	0
MERGE_SYSTEM_PROMPT	merge system prompt with user input, for models not support system role, 1 or 0	0
OLLAMA_CHAT_MODEL	ollama chat model	llama3
OLLAMA_EMBEDDING_MODEL	ollama embedding model	nomic-embed-text
OLLAMA_URL	ollama url	http://localhost:11434
OPENAI_API_BASE	openai compatible api base url	https://api.openai.com/v1
OPENAI_API_KEY	openai api key	EMPTY
OPENAI_CHAT_MODEL	openai chat model	gpt-4o-mini
OPENAI_EMBEDDING_MODEL	openai embedding model	text-embedding-3-small
OTEL_SDK_DISABLED	disable OpenTelemetry, false or true	false
OTEL_SERVICE_NAME	OpenTelemetry service name, also used for Pyroscope application name	chat-search
PYROSCOPE_BASIC_AUTH_PASSWORD	Pyroscope basic auth password
PYROSCOPE_BASIC_AUTH_USERNAME	Pyroscope basic auth username
PYROSCOPE_ENABLED	Enable Pyroscope or not, 1 or 0	1
PYROSCOPE_SERVER_ADDRESS	Pyroscope server address	http://localhost:4040
REDIS_URL	redis url	redis://localhost:6379/
REPHRASE_PROMPT	prompt for rephrase	check config.py
RETRIEVAL_QA_CHAT_SYSTEM_PROMPT	prompt for retrieval	check config.py
RETRIEVER_SEARCH_K	retriever search result number	4
RETRIEVER_SELF_QUERY_EXAMPLES	retriever self query examples as json	check config.py
TEXT_SPLIT_CHUNK_OVERLAP	chunk overlap for text split	200
TEXT_SPLIT_CHUNK_SIZE	chunk size for text split	4000
VECTORSTORE_RETRIEVER_SEARCH_KWARGS	search kwargs for redis vectorstore retriever as json	check config.py
VECTORSTORE_RETRIEVER_SEARCH_TYPE	search type for redis vectorstore retriever	mmr
VECTORSTORE_RETRIEVER_SELF_QUERY	whether to enable vectorstore retriever self query, 1 or 0	1
VECTORSTORE_RETRIEVER_WEIGHT	vectorstore retriever weight	0.5
VERBOSE	enable verbose, 1 or 0	0

Start Ollama (Optional)

Follow Ollama instructions

ollama serve
ollama pull llama3
ollama pull nomic-embed-text

Run on host

Install dependencies

pip install poetry==1.7.1
poetry shell
poetry install

Start dependencies

Start redis

docker compose -f compose.redis.yaml up

Launch LangServe

langchain serve

Visit http://localhost:8000/

Run in Docker

There is a compose.yml file for running the app and all dependencies in containers. Suitable for local end to end testing.

docker compose up --build

Visit http://localhost:8000/

Run in Kubernetes

There is a helm chart for deploying the app in Kubernetes.

Config Helm values

Using Helm

cp values.example.yaml values.yaml

Then update values.yaml accordingly.

Add helm repos:

helm repo add chat-search https://hemslo.github.io/chat-search/
helm repo add redis-stack https://redis-stack.github.io/helm-redis-stack/
helm repo add ollama-helm https://otwld.github.io/ollama-helm/

Install/Upgrade chat-search

helm upgrade -i --wait my-chat-search chat-search/chat-search -f values.yaml

Using Skaffold for local development

skaffold run --port-forward

Ingest data

crawl --sitemap-url $SITEMAP_URL --auth-token $AUTH_TOKEN

Check crawl.yml for web crawling,

Example auto ingest after Github Pages deploy, jekyll.yml.

Architecture

Ingest

Loading

flowchart LR
  A(Crawl) --> |doc| B(/ingest)
  B --> |metadata| C(Redis)
  B --> |doc| D(Text Splitter)
  D --> |docs| E(Embedding Model)
  E --> |docs with embeddings| F(Redis)

Query

Loading

flowchart LR
  A((Request)) --> |messages| B(/chat)
  B --> |messages| C(LLM)
  C --> |question| D(Embedding Model)
  D --> |embeddings| E(Redis)
  E --> |relevant docs| F(LLM)
  B --> |messages|F
  F --> |answer| G((Response))

Deployment

Check cicd.yml for Google Cloud Run deployment, deploy-to-cloud-run.