[Bug]:

[phucph98-AI]

Affected Component

AI Agents (Researcher/Developer/...)

Describe the bug

I am using the latest version of Pentagi with a local LLM.
When I create a new workflow with the prompt: "Scan this website: [http://wordpress.test]", Pentagi seems to be using results or data from a previous workflow run. This is causing the current workflow's results to be confusing and inaccurate.
Could you please clarify if this issue is potentially caused by using a local LLM, or if this is a bug in Pentagi?

Expected behavior:
Each new workflow should start with a clean state and not be influenced by previous, unrelated workflow runs.
Actual behavior:
The new workflow's output is mixed with or influenced by results from a previous workflow, leading to confusing results.

Steps to Reproduce

1. Run a workflow in Pentagi.
2. Create a new workflow.
3. Use the prompt: "Scan this website: http://wordpress.test"
4. Observe that the results appear to include or be influenced by data from the previous workflow run.

System Configuration

# PentAGI Environment Variables

CORS_ORIGINS=http://localhost:*,https://localhost:*

COOKIE_SIGNING_SALT=salt # change this to improve security

# Allow to interact with user while executing tasks
ASK_USER=

## LLM Providers
OPENAI_API_KEY=123
OPENAI_SERVER_URL=https://api.openai.com/v1

ANTHROPIC_API_KEY=
ANTHROPIC_SERVER_URL=https://api.anthropic.com/v1

## Custom LLM provider
LLM_SERVER_URL=http://10.12.181.240:1234/v1
LLM_SERVER_KEY=123
LLM_SERVER_MODEL=qwen/qwen3-30b-a3b
LLM_SERVER_CONFIG_PATH=
LLM_SERVER_LEGACY_REASONING=

## Embedding
EMBEDDING_URL=http://10.12.181.240:3001
EMBEDDING_KEY=
EMBEDDING_MODEL=nomic-embed-text:v1.5
EMBEDDING_PROVIDER=ollama
EMBEDDING_BATCH_SIZE=32

## Summarizer
SUMMARIZER_PRESERVE_LAST=
SUMMARIZER_USE_QA=
SUMMARIZER_SUM_MSG_HUMAN_IN_QA=
SUMMARIZER_LAST_SEC_BYTES=
SUMMARIZER_MAX_BP_BYTES=
SUMMARIZER_MAX_QA_SECTIONS=
SUMMARIZER_MAX_QA_BYTES=
SUMMARIZER_KEEP_QA_SECTIONS=

## Assistant
ASSISTANT_USE_AGENTS=false
ASSISTANT_SUMMARIZER_PRESERVE_LAST=
ASSISTANT_SUMMARIZER_LAST_SEC_BYTES=
ASSISTANT_SUMMARIZER_MAX_BP_BYTES=
ASSISTANT_SUMMARIZER_MAX_QA_SECTIONS=
ASSISTANT_SUMMARIZER_MAX_QA_BYTES=
ASSISTANT_SUMMARIZER_KEEP_QA_SECTIONS=

## HTTP proxy to use it in isolation environment
PROXY_URL=

## Scraper URLs and settings
SCRAPER_PUBLIC_URL=
SCRAPER_PRIVATE_URL=https://someuser:somepass@scraper/
LOCAL_SCRAPER_USERNAME=someuser
LOCAL_SCRAPER_PASSWORD=somepass
LOCAL_SCRAPER_MAX_CONCURRENT_SESSIONS=10

## Web server settings
PUBLIC_URL=https://localhost:8443
STATIC_DIR=
STATIC_URL=
SERVER_PORT=8443
SERVER_HOST=0.0.0.0
SERVER_SSL_CRT=
SERVER_SSL_KEY=
SERVER_USE_SSL=true

## OAuth google
OAUTH_GOOGLE_CLIENT_ID=
OAUTH_GOOGLE_CLIENT_SECRET=

## OAuth github
OAUTH_GITHUB_CLIENT_ID=
OAUTH_GITHUB_CLIENT_SECRET=

## Google search engine API
GOOGLE_API_KEY=
GOOGLE_CX_KEY=

## Traversaal search engine API
TRAVERSAAL_API_KEY=

## Tavily search engine API
TAVILY_API_KEY=

## Perplexity search engine API
PERPLEXITY_API_KEY=
PERPLEXITY_MODEL=
PERPLEXITY_CONTEXT_SIZE=

## Langfuse observability settings
LANGFUSE_BASE_URL=
LANGFUSE_PROJECT_ID=
LANGFUSE_PUBLIC_KEY=
LANGFUSE_SECRET_KEY=

## OpenTelemetry observability settings
OTEL_HOST=

## Docker client settings to run primary terminal container
DOCKER_HOST=unix:///var/run/docker.sock
DOCKER_TLS_VERIFY=
DOCKER_CERT_PATH=

## Docker settings inside primary terminal container
DOCKER_INSIDE=true # enable to use docker socket
DOCKER_NET_ADMIN=true # enable to use net_admin capability
DOCKER_SOCKET=/var/run/docker.sock # path on host machine
DOCKER_NETWORK=pentagi-network # must be exist
DOCKER_PUBLIC_IP=0.0.0.0 # public ip of host machine
DOCKER_WORK_DIR=
DOCKER_DEFAULT_IMAGE=

# Postgres (pgvector) settings
PENTAGI_POSTGRES_USER=postgres
PENTAGI_POSTGRES_PASSWORD=postgres # change this to improve security
PENTAGI_POSTGRES_DB=pentagidb


# Langfuse Environment Variables

## Langfuse Postgres
LANGFUSE_POSTGRES_USER=postgres
LANGFUSE_POSTGRES_PASSWORD=postgres # change this to improve security
LANGFUSE_POSTGRES_DB=langfuse
LANGFUSE_POSTGRES_VERSION=16

## Langfuse Clickhouse
LANGFUSE_CLICKHOUSE_USER=clickhouse
LANGFUSE_CLICKHOUSE_PASSWORD=clickhouse # change this to improve security
LANGFUSE_CLICKHOUSE_URL=http://langfuse-clickhouse:8123
LANGFUSE_CLICKHOUSE_MIGRATION_URL=clickhouse://langfuse-clickhouse:9000
LANGFUSE_CLICKHOUSE_CLUSTER_ENABLED=false

## Langfuse S3
LANGFUSE_S3_BUCKET=langfuse
LANGFUSE_S3_REGION=auto
LANGFUSE_S3_ACCESS_KEY_ID=accesskey # change this to improve security
LANGFUSE_S3_SECRET_ACCESS_KEY=secretkey # change this to improve security
LANGFUSE_S3_ENDPOINT=http://langfuse-minio:9000
LANGFUSE_S3_FORCE_PATH_STYLE=true
LANGFUSE_S3_EVENT_UPLOAD_PREFIX=events/
LANGFUSE_S3_MEDIA_UPLOAD_PREFIX=media/

## Langfuse Redis
LANGFUSE_REDIS_HOST=langfuse-redis
LANGFUSE_REDIS_PORT=6379
LANGFUSE_REDIS_AUTH=redispassword # change this to improve security

## Langfuse web app security settings
LANGFUSE_SALT=salt # change this to improve security
LANGFUSE_ENCRYPTION_KEY=0000000000000000000000000000000000000000000000000000000000000000 # change this to improve security

## Langfuse web app nextauth settings
LANGFUSE_NEXTAUTH_URL=http://localhost:4000
LANGFUSE_NEXTAUTH_SECRET=secret # change this to improve security

## Langfuse extra settings
LANGFUSE_ENABLE_EXPERIMENTAL_FEATURES=true
LANGFUSE_TELEMETRY_ENABLED=false
LANGFUSE_LOG_LEVEL=info

## Langfuse init settings
LANGFUSE_INIT_ORG_ID=ocm47619l0000872mcd2dlbqwb
LANGFUSE_INIT_ORG_NAME=PentAGI Org
LANGFUSE_INIT_PROJECT_ID=cm47619l0000872mcd2dlbqwb
LANGFUSE_INIT_PROJECT_NAME=PentAGI
LANGFUSE_INIT_PROJECT_PUBLIC_KEY=pk-lf-00000000-0000-0000-0000-000000000000 # change this to improve security
LANGFUSE_INIT_PROJECT_SECRET_KEY=sk-lf-00000000-0000-0000-0000-000000000000 # change this to improve security
LANGFUSE_INIT_USER_EMAIL=admin@pentagi.com
LANGFUSE_INIT_USER_NAME=admin
LANGFUSE_INIT_USER_PASSWORD=password # change this to improve security

## Langfuse SDK sync settings
LANGFUSE_SDK_CI_SYNC_PROCESSING_ENABLED=false
LANGFUSE_READ_FROM_POSTGRES_ONLY=false
LANGFUSE_READ_FROM_CLICKHOUSE_ONLY=true
LANGFUSE_RETURN_FROM_CLICKHOUSE=true

## Langfuse license settings
LANGFUSE_EE_LICENSE_KEY=

## Langfuse OpenTelemetry settings
LANGFUSE_OTEL_EXPORTER_OTLP_ENDPOINT=
LANGFUSE_OTEL_SERVICE_NAME=

## Langfuse custom oauth2 settings
LANGFUSE_AUTH_CUSTOM_CLIENT_ID=
LANGFUSE_AUTH_CUSTOM_CLIENT_SECRET=
LANGFUSE_AUTH_CUSTOM_ISSUER=
LANGFUSE_AUTH_CUSTOM_NAME=PentAGI
LANGFUSE_AUTH_CUSTOM_SCOPE=openid email profile
LANGFUSE_AUTH_CUSTOM_CLIENT_AUTH_METHOD=client_secret_post
LANGFUSE_AUTH_CUSTOM_ALLOW_ACCOUNT_LINKING=true

## Langfuse auth settings
LANGFUSE_AUTH_DISABLE_SIGNUP=false # disable signup if PentAGI OAuth2 is used
LANGFUSE_AUTH_SESSION_MAX_AGE=240

## Langfuse allowed organization creators
LANGFUSE_ALLOWED_ORGANIZATION_CREATORS=admin@pentagi.com

## Langfuse default settings for new users
LANGFUSE_DEFAULT_ORG_ID=ocm47619l0000872mcd2dlbqwb
LANGFUSE_DEFAULT_PROJECT_ID=cm47619l0000872mcd2dlbqwb
LANGFUSE_DEFAULT_ORG_ROLE=VIEWER
LANGFUSE_DEFAULT_PROJECT_ROLE=VIEWER

Logs and Artifacts

No response

Screenshots or Recordings

No response

Verification

• I have checked that this issue hasn't been already reported
• I have provided all relevant configuration files (with sensitive data removed)
• I have included relevant logs and error messages
• I am running the latest version of PentAGI

[asdek]

Hello @phucph98-AI

Thank you for the bug report!

There are several ways the previous flow state can be used:

1. Interaction with vector store, where agents write guides for themselves on various topics and inevitably include current IPs or URLs of target applications. However, this shouldn't have a major impact on the process itself, as RAG queries have a prompt that tells how to work with the retrieval results;

2. Long term memory within each individual flow - agents can and should receive information about previously executed actions, but any such queries have the flow ID strictly embedded to avoid exactly the problem you're describing;

3. Short term memory for agents with summarization mode, which works so that a newly launched agent remembers the previous launch context and results according to summarization rules, taking into account SUMMARIZER_* variables. But again, the selection of previous state is limited by flow ID.

Each of these points could potentially contain bugs when using filters, but I can't spot this in the code visually right now and haven't encountered such behavior yet...

To help me progress in understanding where the problem is located, could you please look at the Vector Store tab and try to find clues there?

You need to pay attention to:

• Type of erroneous retrieval - this will be in the title, for example: "Retrieved answer as a tool" or "Retrieved guide about use" or "Retrieved from memory". In these examples, the type would be: "answer", "guide", "memory"
• Task ID and Subtask ID, which are located under the text area. They should belong to the current flow for "memory" type and can cross other flows for all other types

I'll also try to search for the bug in the code. If you can identify examples of such errors, it will greatly speed up the bug fixing process.

For the third point, you need to look at the message chain in Langfuse - we can postpone this for now.