使用 NVIDIA NeMo Retriever、託管 NIM、LanceDB、重新排序與基於事實生成建立多模態 RAG 管線

In this tutorial, we build an advanced multimodal retrieval-augmented generation pipeline with NVIDIA NeMo Retriever.We begin by configuring a Python 3.12 environment, installing the required packages, and performing offline PDF text extraction without relying on a GPU or external API key.

We then extend the workflow with hosted NVIDIA NIM endpoints to detect page elements, extract tables, charts, and infographics, generate dense vector embeddings, and store the processed content in LanceDB.

Finally, we implement dense retrieval, vision-language reranking, metadata-filtered search, grounded response generation with inline citations, and a lightweight recall-at-k evaluation to validate retrieval quality across multimodal document content.

Copy CodeCopiedUse a different Browserimport sys, os, subprocess, textwrap, json, time, warnings warnings.filterwarnings("ignore") assert sys.versioninfo[:2] == (3, 12), ( f"nemo-retriever requires Python 3.12.x (found {sys.version.split()[0]})." "Colab's default runtime is 3.

12; if you changed it, switch back." ) def sh(cmd): print(f"$ {cmd}") subprocess.

run(cmd, shell=True, check=False) try: import nemoretriever print("nemo-retriever already installed") except ImportError: sh("pip install -q --ignore-installed PyJWT nemo-retriever openai") import nemoretriever print("nemo-retriever version:", nemoretriever.

version) from nemoretriever import createingestor try: from nemoretriever.io import tomarkdown, tomarkdownbypage except ImportError: from nemoretriever.common.io import tomarkdown, tomarkdownbypage try: from nemoretriever.retriever import Retriever except ImportError: from nemoretriever.graph.

retriever import Retriever import pandas as pd pd.setoption("display.maxcolwidth", 160) DOC = "multimodaltest.pdf" if not os.path.exists(DOC): sh(f"curl -sL -o {DOC} " "https://raw.githubusercontent.com/NVIDIA/NeMo-Retriever/main/data/multimodaltest.pdf") print("document:", DOC, os.path.

getsize(DOC), "bytes") DOCS = [DOC] print("\n=== STAGE 1: offline text extraction (no API key) ===") offline = ( createingestor(runmode="inprocess", allownogpu=True) .files(DOCS) .

extract( extracttext=True, extracttables=False, extractcharts=False, extractimages=False, extractinfographics=False, usepageelements=False, extractpageasimage=False, method="pdfium", ) ) dfoffline = offline.ingest() print("rows:", dfoffline.shape, "\ncolumns:", list(dfoffline.

columns)) print("\npage 1 text preview:\n", dfoffline.iloc[0]["text"][:400]) We configure the Python 3.12 environment, install NVIDIA NeMo Retriever, and import the required ingestion and retrieval components.We download the sample multimodal PDF and define it as the input document for the pipeline.

We then perform CPU-based offline text extraction with PDFium and inspect the extracted rows, columns, and page content.Copy CodeCopiedUse a different Browserfrom getpass import getpass if not os.environ.get("NVIDIAAPIKEY"): try: from google.colab import userdata os.

environ["NVIDIAAPIKEY"] = userdata.get("NVIDIAAPIKEY") except Exception: os.environ["NVIDIAAPIKEY"] = getpass("NVIDIAAPIKEY (nvapi-...): ").strip() APIKEY = os.environ.get("NVIDIAAPIKEY", "").strip() HAVEKEY = APIKEY.

startswith("nvapi-") print("API key present:", HAVEKEY) PAGEELEMENTSURL = "https://ai.api.nvidia.com/v1/cv/nvidia/nemotron-page-elements-v3" OCRURL = "https://ai.api.nvidia.com/v1/cv/nvidia/nemotron-ocr-v1" TABLESTRUCTURL = "https://ai.api.nvidia.

com/v1/cv/nvidia/nemotron-table-structure-v1" GRAPHICELEMURL = "https://ai.api.nvidia.com/v1/cv/nvidia/nemotron-graphic-elements-v1" EMBEDURL = "https://integrate.api.nvidia.com/v1/embeddings" RERANKURL = "https://ai.api.nvidia.

com/v1/retrieval/nvidia/llama-nemotron-rerank-vl-1b-v2/reranking" CHATURL = "https://integrate.api.nvidia.com/v1" EMBEDMODEL = "nvidia/llama-nemotron-embed-1b-v2" RERANKMODEL = "nvidia/llama-nemotron-rerank-vl-1b-v2" LLMMODEL = "nvidia/llama-3.3-nemotron-super-49b-v1.5" LANCEDBURI, TABLE = ".

/lancedb", "colabdemo" df = dfoffline if HAVEKEY: print("\n=== STAGE 2: multimodal ingest via hosted NIMs ===") ing = ( createingestor( runmode="inprocess", allownogpu=True, errorpolicy="collect", ) .files(DOCS) .

extract( extracttext=True, extracttables=True, extractcharts=True, extractinfographics=True, extractimages=False, method="pdfium", dpi=200, tableoutputformat="markdown", pageelementsinvokeurl=PAGEELEMENTSURL, ocrinvokeurl=OCRURL, tablestructureinvokeurl=TABLESTRUCTURL, graphicelementsinvokeurl=GRAPHICELEMURL, apikey=APIKEY, requesttimeouts=120.

0, splitconfig={"text": {"maxtokens": 512, "overlaptokens": 64}}, ) .dedup(contenthash=True, bboxiou=True, iouthreshold=0.45) .embed( embeddingendpoint=EMBEDURL, modelname=EMBEDMODEL, embedmodelname=EMBEDMODEL, apikey=APIKEY, inputtype="passage", inferencebatchsize=16, nimhttpmaxconcurrent=8, ) .

vdbupload( vdbop="lancedb", vdbkwargs={ "uri": LANCEDBURI, "tablename": TABLE, "overwrite": True, "createindex": True, "indextype": "IVFHNSWSQ", "metric": "l2", }, ) ) t0 = time.time() df = ing.ingest(showprogress=True) print(f"ingested in {time.time()-t0:.1f}s -> {df.

shape}") We securely load the NVIDIA API key and define the hosted NIM endpoints for layout detection, OCR, table extraction, graphic analysis, embedding, reranking, and generation.

We create a multimodal ingestion pipeline that extracts text, tables, charts, and infographics while applying token-aware chunking and content deduplication.We generate embeddings for the extracted content and upload the resulting vectors and metadata to a LanceDB table.

Copy CodeCopiedUse a different Browserprint("\n=== Extraction inspection ===") for col in ["tables", "charts", "infographics", "images"]: if col in df.columns: n = int(df[col].apply(lambda v: len(v) if isinstance(v, (list, tuple)) else 0).

sum()) print(f" {col:<14} {n}") pages = tomarkdownbypage(df) print("\npages rendered to markdown:", list(pages.keys())) print("\n--- page 1 markdown (first 900 chars) ---\n", pages[min(pages)][:900]) fullmd = tomarkdown(df) if fullmd: with open("extracted.md", "w") as f: f.

write(fullmd) print("\nfull document markdown -> extracted.

md") if HAVEKEY: print("\n=== STAGE 3: dense retrieval ===") retriever = Retriever( runmode="service", topk=5, rerank=False, vdbkwargs={"uri": LANCEDBURI, "tablename": TABLE}, embedkwargs={ "embeddingendpoint": EMBEDURL, "modelname": EMBEDMODEL, "embedmodelname": EMBEDMODEL, "apikey": APIKEY, "inputtype": "query", }, ) QUERIES = [ "Given their activities, which animal is responsible for the typos in my documents?

", "What is the most expensive gadget and how much does it cost?", "Which animal is at the beach?", ] def show(hits, label=""): print(f"\n--- {label} ---") for i, h in enumerate(hits, 1): meta = h.get("metadata") if isinstance(meta, str): try: meta = json.

loads(meta) except Exception: meta = {} page = (meta or {}).get("pagenumber", "?") score = h.get("distance", h.get("rerankscore", "")) body = " ".join(str(h.get("text", "")).split())[:180] print(f" {i}.p{page} score={score} {body}") show(retriever.

query(QUERIES[0]), "single query") for q, hits in zip(QUERIES, retriever.queries(QUERIES, topk=3)): show(hits, q[:60]) We inspect the extracted multimodal elements and convert the processed document into page-level and full-document Markdown.

We configure a dense retriever that embeds user queries and searches the LanceDB vector index for the most relevant document chunks.We test both individual and batched queries while displaying page numbers, similarity scores, and retrieved text previews.

Copy CodeCopiedUse a different Browserif HAVEKEY: print("\n=== STAGE 4: retrieve + VL rerank ===") reranking = Retriever( runmode="service", topk=5, rerank=True, vdbkwargs={"uri": LANCEDBURI, "tablename": TABLE}, embedkwargs={ "embeddingendpoint": EMBEDURL, "modelname": EMBED_MODEL, "emb