A conversational RAG agent pipeline using LangGraph
Project description
knowai
An agentic AI pipeline for multiple, large PDF reports interrogation
Set up
- Clone this repository into a local directory of your choosing
- Build a virtual environment
- Install
knowaiby running:pip install .from the root directory of your clone (OR) install usingpip install knowaifrom PyPI. - Configure a
.envfile with the following:AZURE_OPENAI_API_KEY- Your API keyAZURE_OPENAI_ENDPOINT- Your Azure endpointAZURE_OPENAI_DEPLOYMENT- Your LLM deployment name (e.g., "gpt-4o")AZURE_EMBEDDINGS_DEPLOYMENT- Your embeddings model deployment name (e.g., "text-embedding-3-large", defaults to "text-embedding-3-large")AZURE_OPENAI_API_VERSION- Your Azure LLM deployment version (e.g., "2024-02-01")AZURE_OPENAI_EMBEDDINGS_API_VERSION- Your Azure embeddings API version (e.g., "2024-02-01", defaults to "2024-02-01")VECTORSTORE_EMBEDDING_BATCH_SIZE- Optional number of chunks to embed per vectorstore write batch. Defaults to50.
Building the vectorstore
Using the CLI (Recommended)
From the root directory of this repository, run the following from a terminal (ensuring that your virtual environment is active) to build the vectorstore:
python -m knowai.cli_vectorstore build <directory_containing_your_input_pdf_files> --metadata_parquet_path <path_to_metadata.parquet> --vectorstore_path <directory_name_for_vectorstore>
Using the Python API
You can also build vectorstores programmatically:
from knowai import get_retriever_from_directory
retriever = get_retriever_from_directory(
directory_path="path/to/pdfs",
persist_directory="my_vectorstore",
metadata_parquet_path="metadata.parquet",
k=10,
chunk_size=1400,
chunk_overlap=200
)
PDF ingestion and embedding batches
When building a vectorstore, KnowAI processes each PDF page by page. Text is extracted with PyMuPDF, using standard text extraction first and a block-based fallback when a page has no standard text. Page text is split into overlapping chunks, very short chunks are skipped, and each retained chunk is stored as a LangChain Document with the chunk text plus metadata from the parquet file.
Each chunk also gets provenance fields that make retrieval traceable:
file_namesource_pathpagechunk_index
Embedding writes are batched by default to avoid sending oversized inputs to the embeddings model. New chunks are embedded and added to FAISS in batches of up to 50 documents unless VECTORSTORE_EMBEDDING_BATCH_SIZE is set. When updating an existing vectorstore, KnowAI checks existing file_name and page metadata and skips pages already present in the FAISS store.
Inspecting Vectorstores
To inspect an existing vectorstore:
python -m knowai.cli_vectorstore inspect <vectorstore_path>
Or programmatically:
from knowai import load_vectorstore, show_vectorstore_schema, list_vectorstore_files, analyze_vectorstore_chunking
# Load vectorstore
vectorstore = load_vectorstore("my_vectorstore")
# Show schema information
schema = show_vectorstore_schema(vectorstore)
print(f"Total vectors: {schema['total_vectors']}")
print(f"Metadata fields: {schema['metadata_fields']}")
# List files in vectorstore
files = list_vectorstore_files(vectorstore)
print(f"Files: {files}")
# Analyze chunking parameters used to build the vectorstore
analysis = analyze_vectorstore_chunking(vectorstore)
print(f"Estimated chunk size: {analysis['recommended_settings']['chunk_size']}")
print(f"Estimated overlap: {analysis['recommended_settings']['chunk_overlap']}")
By default, this will create a vectorstore using FAISS named "test_faiss_store" in the root directory of your repository.
Running the knowai in a simple chatbot example via streamlit
From the root directory, run the following in a terminal after you have your virtual environment active:
streamlit run app_chat_simple.py
This will open the app in your default browser.
Using knowai
Once your vector store is built, you can use knowai either programmatically or through the provided Streamlit interface.
Python quick‑start
The package ships with the KnowAIAgent class for fully programmatic access
inside notebooks or scripts:
from knowai.core import KnowAIAgent
# Path that you supplied with --vectorstore_path when building
VSTORE_PATH = "test_faiss_store"
agent = KnowAIAgent(vectorstore_path=VSTORE_PATH)
# A single conversational turn
response = await agent.process_turn(
user_question="Summarize the key findings in the 2025 maritime report",
selected_files=["my_report.pdf"],
)
print(response["generation"])
Streaming Responses
For a more responsive user experience, you can enable streaming responses:
def stream_callback(token: str):
"""Called for each token as it's generated."""
print(token, end='', flush=True)
response = await agent.process_turn(
user_question="Summarize the key findings in the 2025 maritime report",
selected_files=["my_report.pdf"],
streaming_callback=stream_callback # Enable streaming
)
The response will be streamed in real-time via the callback, while still being available in the returned dictionary.
The returned dictionary contains:
| Key | Description |
|---|---|
generation |
Final answer synthesised from the selected documents. |
individual_answers |
Per‑file answers (when bypass_individual_gen=False). |
documents_by_file |
Retrieved document chunks keyed by filename. |
raw_documents_for_synthesis |
Raw text block used when bypassing individual generation. |
bypass_individual_generation |
Whether the bypass mode was used for this turn. |
Token Counting Configuration
KnowAI supports two methods for token counting to manage context window limits:
Accurate Token Counting (Default)
- Uses
tiktokenlibrary for precise token estimation - More accurate batch sizing and context management
- Automatically falls back to heuristic method if
tiktokenunavailable
# Default behavior (accurate token counting)
agent = KnowAIAgent(vectorstore_path=VSTORE_PATH)
# Explicit accurate token counting
agent = KnowAIAgent(
vectorstore_path=VSTORE_PATH,
)
Heuristic Token Counting
- Uses character-based estimation (4 characters ≈ 1 token)
- Faster performance, suitable when approximate estimation is sufficient
- Always available as fallback
# Use heuristic token counting
agent = KnowAIAgent(
vectorstore_path=VSTORE_PATH,
)
CLI Configuration
curl -X POST http://127.0.0.1:8000/initialize \
-H "Content-Type: application/json" \
-d '{
"vectorstore_s3_uri": "/path/to/vectorstore",
}'
Benefits of Accurate Token Counting:
- More precise token limits and batch sizing
- Reduced risk of context overflow
- Better resource utilization
- Improved reliability with large document sets
When to Use Heuristic Counting:
tiktokennot available in environment- Performance is critical
- Approximate estimation is sufficient
- Debugging token counting issues
Streamlit chat app
If you prefer a ready‑made UI, launch the demo:
streamlit run app_chat_simple.py
Upload or select PDF files, ask questions in the sidebar, and inspect per‑file answers or the combined response in the main panel.
For advanced configuration options (e.g., conversation history length,
retriever k values, or combine thresholds) see the docstrings in
knowai/core.py and knowai/agent.py.
Containerization
To build and run both the knowai service and the Svelte UI using Docker Compose:
- Ensure Docker and Docker Compose are installed on your machine.
- From the directory containing this README (the repo root), navigate to the Svelte example folder:
cd example_apps/svelte
2a. Compile the Svelte app and package the build as svelte-example:
npm install
npm run build
mv dist svelte-example
- Start the services and build images:
docker compose up --build
This will:- Build the
knowaiservice (listening on port 8000). - Build the
uiservice (Svelte app, listening on port 5173).
- Build the
- Open your browser and visit:
- FastAPI docs: http://localhost:8000/docs
- Svelte UI: http://localhost:5173
- To stop and remove containers, press
CTRL+Cand then run:docker compose down
Running the knowai CLI Locally
You can start the FastAPI micro-service locally without Docker and point it to either a local vectorstore or one hosted on S3.
Using a Local Vectorstore
- Ensure you have a built FAISS vectorstore on disk (e.g.,
test_faiss_store). - Start the service:
python -m knowai.cli
- In another terminal, initialize the session:
curl -X POST http://127.0.0.1:8000/initialize \ -H "Content-Type: application/json" \ -d '{"vectorstore_s3_uri":"/absolute/path/to/your/vectorstore"}'
- Ask a question:
curl -X POST http://127.0.0.1:8000/ask \ -H "Content-Type: application/json" \ -d '{ "session_id":"<session_id>", "question":"Your question here", "selected_files":["file1.pdf","file2.pdf"] }'
Streaming API
For real-time streaming responses, use the /ask-stream endpoint:
curl -X POST http://127.0.0.1:8000/ask-stream \
-H "Content-Type: application/json" \
-d '{
"session_id":"<session_id>",
"question":"Your question here",
"selected_files":["file1.pdf","file2.pdf"]
}' \
--no-buffer
This will stream the response in real-time using Server-Sent Events (SSE). Each token will be sent as it's generated by the LLM.
For more details on streaming functionality, see docs/STREAMING.md.
Using an S3-Hosted Vectorstore
- Start the service:
python -m knowai.cli
- Initialize the session against your S3 bucket:
curl -X POST http://127.0.0.1:8000/initialize \ -H "Content-Type: application/json" \ -d '{"vectorstore_s3_uri":"s3://your-bucket/path"}'
- Ask a question in a similar way:
curl -X POST http://127.0.0.1:8000/ask \ -H "Content-Type: application/json" \ -d '{ "session_id":"<session_id>", "question":"Another question example", "selected_files":[] }'
Enhanced User Feedback
KnowAI provides comprehensive feedback when the search process doesn't find relevant information in your documents.
No-Chunks Feedback
When no text chunks are extracted for a query in a file, KnowAI ensures users are clearly informed:
- Individual File Level: Each file that has no matching content receives a specific message explaining that "The search did not retrieve any document chunks that match your query."
- Synthesis Level: The final response clearly states which files had no relevant content, helping users understand the scope of the search results.
- Progress Tracking: Files with no matching content are tracked separately from files with errors, providing clear distinction in the response.
Example Response
When asking about "climate change impacts" across multiple reports:
I found information about climate change impacts in the following reports:
From report1.pdf (Page 15):
"Global temperatures have increased by 1.1°C since pre-industrial times..."
From report2.pdf (Page 8):
"Sea level rise is accelerating at a rate of 3.3mm per year..."
No matching content found in: report3.pdf (no matching content).
This helps users understand:
- Which files contained relevant information
- Which files were searched but had no matching content
- The specific nature of missing information
Error Handling
KnowAI distinguishes between different types of issues:
- No matching content: Files that were searched but had no relevant chunks
- Content policy violations: Issues with AI provider content filters
- Processing errors: Technical issues during document processing
Each type is handled appropriately and communicated clearly to the user.
Testing
Run the test suite to verify functionality:
# Run all tests
python -m pytest
# Test specific functionality
python -m pytest tests/test_prompts.py -v
python -m pytest tests/test_agent.py -v
python -m pytest tests/test_vectorstore.py -v
# Test no-chunks feedback improvements
python scripts/test_no_chunks_feedback.py
The vectorstore tests cover default embedding batching, VECTORSTORE_EMBEDDING_BATCH_SIZE overrides, incremental updates to existing FAISS stores, and metadata-aware vectorstore inspection utilities.
Contributing
Contributions are welcome! Please feel free to submit a Pull Request.
Individual File Processing
KnowAI supports two processing modes for handling multiple files:
Traditional Batch Processing (Default)
- All documents from all files are combined and processed together
- Faster processing, good for related content across files
Individual File Processing
- Each file is processed separately by the LLM (in parallel, max 10 concurrent), then responses are consolidated
- Ensures each file gets equal attention, better for distinct topics
- Significantly faster than sequential processing for multiple files
- Concurrency limit prevents overwhelming the LLM service
# Enable individual file processing
agent = KnowAIAgent(
vectorstore_path=VSTORE_PATH,
process_files_individually=True # Enable individual processing
)
# Or enable per-request
response = await agent.process_turn(
user_question="What are the main strategies?",
selected_files=["file1.pdf", "file2.pdf"],
process_files_individually=True # Override for this request
)
When to Use Individual File Processing:
- Files contain distinct topics that should be analyzed separately
- You want to ensure each file gets equal attention from the LLM
- You want to see how each file contributes to the final answer
- Dealing with large files that might benefit from focused analysis
For more details, see Individual File Processing Documentation.
Token Counting Configuration
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file knowai-0.6.0.tar.gz.
File metadata
- Download URL: knowai-0.6.0.tar.gz
- Upload date:
- Size: 45.6 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.1.0 CPython/3.13.3
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
12bb213027ed370e0560bfc73ea77b3509ae1919621949d410eaf0c191685532
|
|
| MD5 |
92274181089f5508f81db5c65a22ec83
|
|
| BLAKE2b-256 |
c6a7011c5668a7b50c9c47526cf0e7133ca26887aeacc603364aacb6b10fa1a3
|
File details
Details for the file knowai-0.6.0-py3-none-any.whl.
File metadata
- Download URL: knowai-0.6.0-py3-none-any.whl
- Upload date:
- Size: 49.4 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.1.0 CPython/3.13.3
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
0fc09b99d206b1654cc627ed5b60aa83a605e15c91fbbd9abe9fb5a9df9cfb6d
|
|
| MD5 |
7393c9e83cff080d56a8ce5d7c52814f
|
|
| BLAKE2b-256 |
b3fd15992908275ed66b83af6d11f9f8cf6d642914a4c3d41eb3c4f1d736c9f1
|